Fix post-transcribe freeze: batch embeddings off the main thread + guard load

Two reported symptoms — the transcript screen sitting on its loading spinner
after a transcription, and a follow-up transcription stuck at 0% — both trace
to the same cause: after a transcript saves, embedOne() embeds the whole
lecture in ONE synchronous WASM call on the MAIN thread. Measured in a browser
harness: embedding 150 segments blocks the thread for ~1.86s straight (a 50ms
heartbeat fired once instead of ~35x); on a phone or a long lecture that's many
seconds of total UI freeze. During it the just-navigated transcript screen
can't run its get() callback (looks "stuck loading") and a concurrent
transcription's progress UI is frozen at 0%.

- embeddingStore.embedTranscript: embed in batches of 16, yielding to the event
  loop between batches, so the UI stays responsive. Per-text vectors are
  identical (e5 pools each text independently); this only changes scheduling.
- transcript/[id].tsx: the get() load had NO error handling, so a rejected load
  (or a long freeze) left the spinner up forever. Add .catch -> show a real
  "couldn't load" message instead of an infinite spinner; also catch the
  (non-fatal) media-url load.

Ruled out (via harness) the theory that the e5 embedding model hits the same
WASM MatMulNBits load failure as Whisper — e5 (encoder-only) loads fine at both
q8 and fp32 on WASM, so its dtype is left unchanged.

Proper long-term fix (separate, larger): run embeddings in a Web Worker.

Validated: tsc clean, 282 tests, web export OK.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.gitea/workflows/ci.yml

@@ -7,12 +7,12 @@
 # `docker compose` simply builds and (re)creates the `wisp` container on the host.
 #
 # Triggers:
-#   push to master -> test, then deploy-web
+#   push to master -> test, then deploy-web, then build-apk
 #   pull_request   -> test only
 #
-# Required repo secrets (Settings -> Actions -> Secrets) — used by the APK job
+# Required repo secrets (Settings -> Actions -> Secrets) — used by the build-apk
-# once it is enabled (see TODO at the bottom): ANDROID_KEYSTORE_BASE64,
+# job: ANDROID_KEYSTORE_BASE64, ANDROID_KEYSTORE_PASSWORD, ANDROID_KEY_ALIAS,
-# ANDROID_KEYSTORE_PASSWORD, ANDROID_KEY_ALIAS, ANDROID_KEY_PASSWORD.
+# ANDROID_KEY_PASSWORD.
 # ---------------------------------------------------------------------------
 
 name: CI
@@ -20,6 +20,7 @@ name: CI
 on:
   push:
     branches: [master]
+    tags: ['v*']
   pull_request:
 
 concurrency:
@@ -67,13 +68,19 @@ jobs:
   # ---- build-apk: signed release APK, served at /wisp.apk ------------------
   # Socket-safe: the whole Android build runs inside `docker buildx build`
   # (context streamed to the host BuildKit — no host bind-mounts), signed via
-  # BuildKit --secret from the ANDROID_* repo secrets. We then extract the APK
+  # BuildKit --secret from the ANDROID_* repo secrets. The `apk` target +
-  # with docker create/cp and copy it into the running `wisp` web container so
+  # `--output type=local` write just the APK to ./out (no image load), which we
-  # it is downloadable at https://wisp.briggen.dev/wisp.apk.
+  # stream into /srv/wisp/wisp.apk so it's downloadable at
+  # https://wisp.briggen.dev/wisp.apk. Speed comes from persistent BuildKit
+  # cache mounts (ccache for C++, Gradle build cache) declared in the Dockerfile.
   build-apk:
+    # Auto-build the APK on every push to master (arm64-only keeps it ~minutes).
+    # Runs after deploy-web so the web updates promptly, then the APK follows and
+    # is published to /srv/wisp/wisp.apk (served at /wisp.apk via the compose mount).
     needs: [test, deploy-web]
     if: github.event_name == 'push' && github.ref == 'refs/heads/master'
     runs-on: ubuntu-latest
+    timeout-minutes: 120
     steps:
       - uses: actions/checkout@v4
       - name: Stage signing secrets as files
@@ -84,21 +91,26 @@ jobs:
           printf '%s' "${{ secrets.ANDROID_KEYSTORE_PASSWORD }}" > "$RUNNER_TEMP/sec/ks_pass"
           printf '%s' "${{ secrets.ANDROID_KEY_ALIAS }}"         > "$RUNNER_TEMP/sec/ks_alias"
           printf '%s' "${{ secrets.ANDROID_KEY_PASSWORD }}"      > "$RUNNER_TEMP/sec/ks_keypass"
-      - name: Build signed APK image
+      - name: Build + extract the signed APK (export-only target, no image load)
         run: |
+          # `--target apk --output type=local` writes ONLY app-release.apk to
+          # ./out — no `--load` of the ~15GB build image (saves the multi-minute
+          # export/unpack each run and stops that image filling the host disk).
+          # The persistent ccache + Gradle build cache (BuildKit cache mounts in
+          # the Dockerfile) make C++ recompiles near-instant on warm runs.
           docker buildx build \
             --secret id=ks_b64,src="$RUNNER_TEMP/sec/ks_b64" \
             --secret id=ks_pass,src="$RUNNER_TEMP/sec/ks_pass" \
             --secret id=ks_alias,src="$RUNNER_TEMP/sec/ks_alias" \
             --secret id=ks_keypass,src="$RUNNER_TEMP/sec/ks_keypass" \
-            -f docker/android.Dockerfile -t wisp-android:ci --load .
+            --target apk --output type=local,dest=./out \
-      - name: Extract + publish APK to /wisp.apk
+            -f docker/android.Dockerfile .
+          ls -lh ./out/app-release.apk
+      - name: Publish APK to the host (served at /wisp.apk via the compose volume)
         run: |
-          cid=$(docker create wisp-android:ci)
+          # Stream the APK into a host-mounted dir so it survives web container
-          docker cp "$cid:/workspace/android/app/build/outputs/apk/release/app-release.apk" ./app-release.apk
+          # recreation (the wisp container mounts /srv/wisp/wisp.apk read-only).
-          docker rm "$cid"
+          docker run --rm -i -v /srv/wisp:/out alpine:3 sh -c 'mkdir -p /out && cat > /out/wisp.apk' < ./out/app-release.apk
-          ls -lh app-release.apk
-          docker cp ./app-release.apk wisp:/usr/share/nginx/html/wisp.apk
       - name: Cleanup staged secrets
         if: always()
         run: rm -rf "$RUNNER_TEMP/sec"
@@ -107,4 +119,4 @@ jobs:
         continue-on-error: true
         with:
           name: wisp-release-apk
-          path: app-release.apk
+          path: ./out/app-release.apk

deploy/wisp.compose.yml

@@ -27,6 +27,12 @@ services:
       - "traefik.http.routers.wisp.entrypoints=websecure"
       - "traefik.http.routers.wisp.tls.certresolver=letsencrypt"
       - "traefik.http.services.wisp.loadbalancer.server.port=8080"
+    # The downloadable Android APK, published to the host by the build-apk CI job
+    # (or a local build). Mounted read-only so it survives web container rebuilds.
+    # NOTE: /srv/wisp/wisp.apk must exist on the host before `up` (else Docker
+    # creates a directory there). The publish step creates it.
+    volumes:
+      - /srv/wisp/wisp.apk:/usr/share/nginx/html/wisp.apk:ro
     networks:
       - proxy
 

docker/android.Dockerfile

@@ -66,14 +66,53 @@ RUN yes | sdkmanager --licenses > /dev/null \
         "cmake;${ANDROID_CMAKE_VERSION}" > /dev/null \
     && yes | sdkmanager --licenses > /dev/null
 
+# Build accelerators, installed AFTER the (large, slow) SDK/NDK layer so that
+# layer stays cached:
+#   - ccache : short-circuits recompiling unchanged C++ across builds. The bulk
+#              of the build is third-party native modules (whisper.cpp via
+#              whisper.rn, reanimated/worklets, gesture-handler); see the NDK
+#              toolchain patch below for how it's wired in globally.
+#   - nodejs : scripts/ci-android-sign.sh runs a small Node patch script (this
+#              image ships Bun, not Node). Lives here so it's cached, not
+#              reinstalled on every source change.
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends ccache nodejs \
+    && rm -rf /var/lib/apt/lists/*
+
+# Route EVERY NDK CMake C/C++ compile + link through ccache. AGP invokes each
+# externalNativeBuild project with this NDK toolchain file, so setting the
+# compiler launcher here applies ccache to all native modules at once — no need
+# to patch each module's CMakeLists (third-party ones don't honor ccache on
+# their own). Guarded on ccache being present; a no-op otherwise.
+RUN printf '\n%s\n' \
+    '# --- WISP: route NDK CMake compiles/links through ccache ---' \
+    'find_program(WISP_CCACHE_PROGRAM ccache)' \
+    'if(WISP_CCACHE_PROGRAM)' \
+    '  set(CMAKE_C_COMPILER_LAUNCHER   "${WISP_CCACHE_PROGRAM}")' \
+    '  set(CMAKE_CXX_COMPILER_LAUNCHER "${WISP_CCACHE_PROGRAM}")' \
+    'endif()' \
+    >> "${ANDROID_HOME}/ndk/${ANDROID_NDK_VERSION}/build/cmake/android.toolchain.cmake"
+
+# ccache tuning. CCACHE_DIR is mounted as a persistent BuildKit cache on the
+# gradle RUN below, so hits survive across builds. COMPILERCHECK=content keeps
+# hits valid across image rebuilds; the sloppiness flags raise the hit rate for
+# RN's headers without risking incorrect cache hits for release builds.
+ENV CCACHE_DIR=/root/.ccache \
+    CCACHE_MAXSIZE=5G \
+    CCACHE_COMPILERCHECK=content \
+    CCACHE_SLOPPINESS=time_macros,include_file_mtime,include_file_ctime,file_macro \
+    CCACHE_BASEDIR=/workspace
+
 WORKDIR /workspace
 
 # ---------------------------------------------------------------------------
 FROM toolchain AS build
 
-# Install JS deps first (cached unless package.json/bun.lock change).
+# Install JS deps first (cached unless package.json/bun.lock change). The bun
+# cache mount keeps re-installs fast even when the lockfile DOES change.
 COPY package.json bun.lock ./
-RUN bun install --frozen-lockfile
+RUN --mount=type=cache,target=/root/.bun/install/cache \
+    bun install --frozen-lockfile
 
 # Bring in the rest of the source (.dockerignore excludes node_modules, android/, ios/, dist...).
 COPY . .
@@ -81,23 +120,44 @@ COPY . .
 # Generate the native android/ project (autolinks whisper.rn, reanimated, etc.).
 RUN bunx expo prebuild --platform android --no-install
 
-# Node is required by scripts/ci-android-sign.sh (this image ships Bun, not Node).
-# Installed AFTER prebuild so the cached toolchain + prebuild layers are reused.
-RUN apt-get update && apt-get install -y --no-install-recommends nodejs && rm -rf /var/lib/apt/lists/*
-
 # Decode the keystore + wire release signing, then assemble the signed APK.
 # Secrets are mounted as files under /run/secrets and exported only for this RUN.
+#
+# Two persistent BuildKit cache mounts make rebuilds fast (they survive the
+# `COPY . .` layer invalidation that happens on every source change):
+#   - /root/.ccache : compiled C++ objects (the big win — whisper.cpp etc.).
+#   - /opt/gradle   : GRADLE_USER_HOME — the Gradle build cache (--build-cache),
+#                     resolved dependencies, and the downloaded wrapper dist.
+# --build-cache + --parallel let unchanged Kotlin/Java/resource/dex tasks be
+# served from cache and independent module tasks run concurrently.
 RUN --mount=type=secret,id=ks_b64 \
     --mount=type=secret,id=ks_pass \
     --mount=type=secret,id=ks_alias \
     --mount=type=secret,id=ks_keypass \
+    --mount=type=cache,target=/root/.ccache \
+    --mount=type=cache,target=/opt/gradle \
     ANDROID_KEYSTORE_BASE64="$(cat /run/secrets/ks_b64)" \
     ANDROID_KEYSTORE_PASSWORD="$(cat /run/secrets/ks_pass)" \
     ANDROID_KEY_ALIAS="$(cat /run/secrets/ks_alias)" \
     ANDROID_KEY_PASSWORD="$(cat /run/secrets/ks_keypass)" \
     bash scripts/ci-android-sign.sh \
     && cd android \
-    && ./gradlew assembleRelease --no-daemon -x lint --stacktrace
+    && ./gradlew assembleRelease \
+        -PreactNativeArchitectures=arm64-v8a \
+        --build-cache --parallel \
+        --no-daemon --max-workers=2 -x lint --stacktrace \
+    && ccache --show-stats --verbose 2>/dev/null | grep -iE 'cacheable|hit|miss|cache size' || true
 
-# The signed APK (handy for `docker create` + `docker cp` extraction in CI).
+# The signed APK lands at:
-# /workspace/android/app/build/outputs/apk/release/app-release.apk
+#   /workspace/android/app/build/outputs/apk/release/app-release.apk
+
+# ---------------------------------------------------------------------------
+# apk : a tiny export-only stage holding JUST the signed APK.
+#
+# CI builds this stage with `--target apk --output type=local,dest=...`, which
+# writes only app-release.apk to the host — NO `--load` of the ~15GB build image
+# (saves the multi-minute image export/unpack on every push AND avoids that
+# image piling up on the server's disk). The COPY pulls from `build`, so the
+# whole build still runs; only the output is slimmed.
+FROM scratch AS apk
+COPY --from=build /workspace/android/app/build/outputs/apk/release/app-release.apk /app-release.apk

eslint.config.js

@@ -0,0 +1,10 @@
+// https://docs.expo.dev/guides/using-eslint/
+const { defineConfig } = require('eslint/config');
+const expoConfig = require("eslint-config-expo/flat");
+
+module.exports = defineConfig([
+  expoConfig,
+  {
+    ignores: ["dist/*"],
+  }
+]);

package.json

@@ -16,6 +16,7 @@
     "expo-image": "~56.0.11",
     "expo-linking": "~56.0.14",
     "expo-router": "~56.2.10",
+    "expo-sharing": "~56",
     "expo-splash-screen": "~56.0.10",
     "expo-sqlite": "^56.0.5",
     "expo-status-bar": "~56.0.4",
@@ -25,6 +26,7 @@
     "react": "19.2.3",
     "react-dom": "19.2.3",
     "react-native": "0.85.3",
+    "react-native-audio-api": "0.12.2",
     "react-native-gesture-handler": "~2.31.1",
     "react-native-reanimated": "4.3.1",
     "react-native-safe-area-context": "~5.7.0",
@@ -38,6 +40,8 @@
   "devDependencies": {
     "@types/node": "^25.9.3",
     "@types/react": "~19.2.2",
+    "eslint": "^9.0.0",
+    "eslint-config-expo": "~56.0.4",
     "fake-indexeddb": "^6.2.5",
     "fast-check": "^4.8.0",
     "typescript": "~6.0.3",

src/app/(tabs)/_layout.tsx

@@ -0,0 +1,51 @@
+// Bottom tab navigation for the primary destinations (mobile-first). Secondary
+// screens (Record, Transcript, Courses, Quiz, Bibliography) are pushed on the
+// root Stack from within these tabs. Route-group parens keep URLs unchanged
+// (/, /search, /study, /ask, /settings).
+import { Tabs } from 'expo-router';
+import { Text, type ColorValue } from 'react-native';
+
+import { useTheme } from '@/hooks/use-theme';
+
+const ACCENT = '#3c87f7';
+
+function TabIcon({ glyph, color }: { glyph: string; color: ColorValue }) {
+  return <Text style={{ fontSize: 20, color }}>{glyph}</Text>;
+}
+
+export default function TabsLayout() {
+  const theme = useTheme();
+  return (
+    <Tabs
+      screenOptions={{
+        tabBarActiveTintColor: ACCENT,
+        tabBarInactiveTintColor: theme.textSecondary,
+        tabBarStyle: { backgroundColor: theme.background, borderTopColor: theme.backgroundElement },
+        headerStyle: { backgroundColor: theme.background },
+        headerTitleStyle: { color: theme.text },
+        headerTintColor: theme.text,
+        headerShadowVisible: false,
+      }}>
+      <Tabs.Screen
+        name="index"
+        options={{ title: 'Wisp', tabBarLabel: 'Library', tabBarIcon: ({ color }) => <TabIcon glyph="📚" color={color} /> }}
+      />
+      <Tabs.Screen
+        name="search"
+        options={{ title: 'Search', tabBarIcon: ({ color }) => <TabIcon glyph="🔎" color={color} /> }}
+      />
+      <Tabs.Screen
+        name="study"
+        options={{ title: 'Study', tabBarIcon: ({ color }) => <TabIcon glyph="🎴" color={color} /> }}
+      />
+      <Tabs.Screen
+        name="ask"
+        options={{ title: 'Ask', tabBarIcon: ({ color }) => <TabIcon glyph="✨" color={color} /> }}
+      />
+      <Tabs.Screen
+        name="settings"
+        options={{ title: 'Settings', tabBarIcon: ({ color }) => <TabIcon glyph="⚙️" color={color} /> }}
+      />
+    </Tabs>
+  );
+}

src/app/(tabs)/ask.tsx

@@ -0,0 +1,314 @@
+import { useFocusEffect, useRouter } from 'expo-router';
+import { useCallback, useState } from 'react';
+import {
+  ActivityIndicator,
+  Pressable,
+  ScrollView,
+  StyleSheet,
+  TextInput,
+  View,
+} from 'react-native';
+
+import { ThemedText } from '@/components/themed-text';
+import { ThemedView } from '@/components/themed-view';
+import { MaxContentWidth, Spacing } from '@/constants/theme';
+import { useTheme } from '@/hooks/use-theme';
+import type { RagAnswer, RagCitation } from '@/lib/generation/engine';
+import { formatClock } from '@/lib/format';
+import { useCourses } from '@/stores/coursesStore';
+import { useAi } from '@/stores/aiStore';
+import { useEmbedding } from '@/stores/embeddingStore';
+
+const ACCENT = '#3c87f7';
+
+// Course filter is a course id, or `null` for "all courses".
+type CourseSel = string | null;
+
+export default function AskScreen() {
+  const theme = useTheme();
+  const router = useRouter();
+
+  const courses = useCourses((s) => s.items);
+  const refreshCourses = useCourses((s) => s.refresh);
+
+  // Whether any lectures have been indexed at all (used to nudge the user to
+  // build the search index when an empty/ungrounded answer comes back).
+  const pending = useEmbedding((s) => s.pending);
+  const refreshPending = useEmbedding((s) => s.refreshPending);
+
+  // AI engine state (model download / readiness) is observed from the store so
+  // the loading bar reflects an in-flight on-device model download.
+  const modelStatus = useAi((s) => s.modelStatus);
+  const progress = useAi((s) => s.progress);
+
+  const [question, setQuestion] = useState('');
+  const [courseId, setCourseId] = useState<CourseSel>(null);
+  const [answering, setAnswering] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+  const [result, setResult] = useState<RagAnswer | null>(null);
+
+  useFocusEffect(
+    useCallback(() => {
+      void refreshCourses();
+      void refreshPending();
+    }, [refreshCourses, refreshPending]),
+  );
+
+  const onAsk = useCallback(async () => {
+    const q = question.trim();
+    if (!q || answering) return;
+    setAnswering(true);
+    setError(null);
+    setResult(null);
+    try {
+      const ans = await useAi.getState().ask(q, { courseId });
+      setResult(ans);
+    } catch (e) {
+      setError(e instanceof Error ? e.message : 'Something went wrong.');
+    } finally {
+      setAnswering(false);
+    }
+  }, [question, courseId, answering]);
+
+  const openCitation = (c: RagCitation) =>
+    router.push({
+      pathname: '/transcript/[id]',
+      params: { id: c.transcriptId, t: String(Math.floor(c.start)) },
+    });
+
+  // The model is downloading/preparing on-device (WebLLM). Show a progress bar.
+  const loadingModel = modelStatus === 'loading';
+
+  return (
+    <ThemedView style={styles.fill}>
+      <ScrollView contentContainerStyle={styles.content} keyboardShouldPersistTaps="handled">
+        <ThemedText type="small" themeColor="textSecondary">
+          Ask a question and get an answer grounded in your lectures — every claim links back to the
+          moment it came from.
+        </ThemedText>
+
+        {courses.length > 0 && (
+          <ScrollView
+            horizontal
+            showsHorizontalScrollIndicator={false}
+            contentContainerStyle={styles.filterBar}>
+            <FilterChip label="All courses" active={courseId === null} onPress={() => setCourseId(null)} />
+            {courses.map((c) => (
+              <FilterChip
+                key={c.id}
+                label={c.name}
+                active={courseId === c.id}
+                onPress={() => setCourseId(c.id)}
+              />
+            ))}
+          </ScrollView>
+        )}
+
+        <TextInput
+          value={question}
+          onChangeText={setQuestion}
+          onSubmitEditing={() => void onAsk()}
+          returnKeyType="search"
+          autoFocus
+          multiline
+          placeholder="Ask your lectures anything…"
+          placeholderTextColor={theme.textSecondary}
+          style={[styles.input, { color: theme.text, backgroundColor: theme.backgroundElement }]}
+        />
+
+        <Pressable
+          onPress={() => void onAsk()}
+          disabled={answering || question.trim() === ''}
+          style={({ pressed }) => [
+            styles.askBtn,
+            { opacity: answering || question.trim() === '' ? 0.5 : pressed ? 0.85 : 1 },
+          ]}>
+          <ThemedText style={styles.askBtnText}>Ask</ThemedText>
+        </Pressable>
+
+        {loadingModel && (
+          <ThemedView type="backgroundElement" style={styles.card}>
+            <ThemedText type="smallBold">Preparing on-device AI… {Math.round(progress * 100)}%</ThemedText>
+            <ProgressBar value={progress} />
+            <ThemedText type="small" themeColor="textSecondary">
+              The model downloads once, then runs locally on your device.
+            </ThemedText>
+          </ThemedView>
+        )}
+
+        {answering && !loadingModel && (
+          <View style={styles.center}>
+            <ActivityIndicator />
+            <ThemedText type="small" themeColor="textSecondary" style={styles.centerText}>
+              Thinking…
+            </ThemedText>
+          </View>
+        )}
+
+        {error && (
+          <ThemedView type="backgroundElement" style={styles.card}>
+            <ThemedText type="smallBold">Couldn&apos;t answer that</ThemedText>
+            <ThemedText type="small" themeColor="textSecondary">
+              {error}
+            </ThemedText>
+          </ThemedView>
+        )}
+
+        {result && !answering && (
+          <Answer
+            result={result}
+            indexEmpty={pending > 0}
+            onOpenCitation={openCitation}
+          />
+        )}
+      </ScrollView>
+    </ThemedView>
+  );
+}
+
+function Answer({
+  result,
+  indexEmpty,
+  onOpenCitation,
+}: {
+  result: RagAnswer;
+  indexEmpty: boolean;
+  onOpenCitation: (c: RagCitation) => void;
+}) {
+  // Nothing relevant was retrieved — we refuse to invent an answer.
+  if (!result.grounded) {
+    return (
+      <ThemedView type="backgroundElement" style={styles.card}>
+        <ThemedText type="smallBold">
+          I couldn&apos;t find anything about that in your lectures.
+        </ThemedText>
+        {indexEmpty && (
+          <ThemedText type="small" themeColor="textSecondary">
+            Tip: build the search index on the Search screen so your lectures become searchable.
+          </ThemedText>
+        )}
+      </ThemedView>
+    );
+  }
+
+  return (
+    <View style={styles.answerWrap}>
+      {result.generated ? (
+        <ThemedView type="backgroundElement" style={styles.answerCard}>
+          <ThemedText type="default">{result.answer}</ThemedText>
+        </ThemedView>
+      ) : (
+        // Grounded but no LLM available: deterministic search-only fallback.
+        <ThemedView type="backgroundElement" style={styles.card}>
+          <ThemedText type="small" themeColor="textSecondary">
+            On-device AI needs WebGPU, or add an API key in Settings — here are the matching moments:
+          </ThemedText>
+        </ThemedView>
+      )}
+
+      <ThemedText type="smallBold" style={styles.sourcesHeading}>
+        Sources
+      </ThemedText>
+      {result.citations.map((c, i) => (
+        <CitationChip
+          key={`${c.transcriptId}:${c.segmentId ?? i}`}
+          citation={c}
+          onPress={() => onOpenCitation(c)}
+        />
+      ))}
+
+      <ThemedText type="small" themeColor="textSecondary" style={styles.disclaimer}>
+        AI answers can be wrong — every claim links to the lecture; verify against the source.
+      </ThemedText>
+    </View>
+  );
+}
+
+function CitationChip({
+  citation,
+  onPress,
+}: {
+  citation: RagCitation;
+  onPress: () => void;
+}) {
+  return (
+    <Pressable onPress={onPress} style={({ pressed }) => [pressed && styles.pressed]}>
+      <ThemedView type="backgroundElement" style={styles.card}>
+        <ThemedText type="small" numberOfLines={3}>
+          {citation.text}
+        </ThemedText>
+        <ThemedText type="small" style={styles.citationTime}>
+          ({formatClock(citation.start)})
+        </ThemedText>
+      </ThemedView>
+    </Pressable>
+  );
+}
+
+function FilterChip({
+  label,
+  active,
+  onPress,
+}: {
+  label: string;
+  active: boolean;
+  onPress: () => void;
+}) {
+  const theme = useTheme();
+  return (
+    <Pressable
+      onPress={onPress}
+      style={[styles.filterChip, { backgroundColor: active ? ACCENT : theme.backgroundElement }]}>
+      <ThemedText type="small" style={active ? styles.chipActive : undefined}>
+        {label}
+      </ThemedText>
+    </Pressable>
+  );
+}
+
+function ProgressBar({ value }: { value: number }) {
+  return (
+    <View style={styles.track}>
+      <View style={[styles.bar, { width: `${Math.max(2, Math.min(100, value * 100))}%` }]} />
+    </View>
+  );
+}
+
+const styles = StyleSheet.create({
+  fill: { flex: 1 },
+  content: {
+    padding: Spacing.three,
+    gap: Spacing.three,
+    maxWidth: MaxContentWidth,
+    width: '100%',
+    alignSelf: 'center',
+  },
+  filterBar: { gap: Spacing.two, paddingVertical: Spacing.one, paddingRight: Spacing.three },
+  filterChip: { paddingHorizontal: Spacing.three, paddingVertical: Spacing.one, borderRadius: 999 },
+  chipActive: { color: '#fff', fontWeight: '700' },
+  input: {
+    borderRadius: Spacing.two,
+    paddingHorizontal: Spacing.three,
+    paddingVertical: Spacing.two,
+    fontSize: 15,
+    minHeight: 48,
+  },
+  askBtn: {
+    backgroundColor: ACCENT,
+    paddingVertical: Spacing.three,
+    borderRadius: Spacing.three,
+    alignItems: 'center',
+  },
+  askBtnText: { color: '#fff', fontWeight: '700', fontSize: 16 },
+  card: { padding: Spacing.three, borderRadius: Spacing.three, gap: Spacing.two },
+  answerWrap: { gap: Spacing.three },
+  answerCard: { padding: Spacing.three, borderRadius: Spacing.three, gap: Spacing.two },
+  sourcesHeading: { marginTop: Spacing.one },
+  citationTime: { color: ACCENT, fontWeight: '700' },
+  disclaimer: { marginTop: Spacing.one, fontStyle: 'italic' },
+  center: { alignItems: 'center', gap: Spacing.two, paddingVertical: Spacing.four },
+  centerText: { textAlign: 'center' },
+  track: { height: 6, borderRadius: 3, backgroundColor: '#88888833', overflow: 'hidden' },
+  bar: { height: 6, borderRadius: 3, backgroundColor: ACCENT },
+  pressed: { opacity: 0.7 },
+});

src/app/(tabs)/index.tsx

@@ -0,0 +1,264 @@
+import { Link, useFocusEffect, useRouter } from 'expo-router';
+import { useCallback, useState } from 'react';
+import {
+  ActivityIndicator,
+  Pressable,
+  ScrollView,
+  StyleSheet,
+  TextInput,
+  View,
+} from 'react-native';
+
+import { PreCaptureSheet, type PreCaptureResult } from '@/components/PreCaptureSheet';
+import { ThemedText } from '@/components/themed-text';
+import { ThemedView } from '@/components/themed-view';
+import { MaxContentWidth, Spacing } from '@/constants/theme';
+import { useTheme } from '@/hooks/use-theme';
+import type { TranscriptMeta } from '@/lib/db';
+import { formatClock } from '@/lib/format';
+import { pickAudio, type PickedAudio } from '@/lib/pickAudio';
+import { useCourses } from '@/stores/coursesStore';
+import { useTranscribe } from '@/stores/transcribeStore';
+import { useTranscripts } from '@/stores/transcriptsStore';
+
+export default function LibraryScreen() {
+  const theme = useTheme();
+  const router = useRouter();
+  const { items, loading, query, setQuery, refresh, remove, courseFilter, setCourseFilter } =
+    useTranscripts();
+  const courses = useCourses((s) => s.items);
+  const refreshCourses = useCourses((s) => s.refresh);
+  const job = useTranscribe();
+  const [picked, setPicked] = useState<PickedAudio | null>(null);
+
+  useFocusEffect(
+    useCallback(() => {
+      void refresh();
+      void refreshCourses();
+    }, [refresh, refreshCourses]),
+  );
+
+  const busy = job.status === 'loading' || job.status === 'transcribing';
+
+  const onNew = useCallback(async () => {
+    if (busy) return;
+    const p = await pickAudio();
+    if (p) setPicked(p);
+  }, [busy]);
+
+  const onStart = useCallback(
+    async (r: PreCaptureResult) => {
+      const p = picked;
+      setPicked(null);
+      if (!p) return;
+      const id = await useTranscribe.getState().start(p, r);
+      if (id) router.push({ pathname: '/transcript/[id]', params: { id } });
+    },
+    [picked, router],
+  );
+
+  const courseName = (cid?: string | null) =>
+    cid ? courses.find((c) => c.id === cid)?.name ?? 'Course' : null;
+
+  return (
+    <ThemedView style={styles.fill}>
+      <ScrollView contentContainerStyle={styles.content}>
+        <View style={styles.subHeader}>
+          <ThemedText type="small" themeColor="textSecondary" style={styles.flex}>
+            On your device — nothing uploaded.
+          </ThemedText>
+          <Link href="/courses" asChild>
+            <Pressable hitSlop={8}>
+              <ThemedText type="link" themeColor="textSecondary">Courses ›</ThemedText>
+            </Pressable>
+          </Link>
+        </View>
+
+        <View style={styles.captureRow}>
+          <Link href="/record" asChild>
+            <Pressable
+              disabled={busy}
+              style={({ pressed }) => [
+                styles.captureBtn,
+                { backgroundColor: '#e5484d', opacity: busy ? 0.5 : pressed ? 0.85 : 1 },
+              ]}>
+              <ThemedText style={styles.captureText}>🎙 Record</ThemedText>
+            </Pressable>
+          </Link>
+          <Pressable
+            onPress={onNew}
+            disabled={busy}
+            style={({ pressed }) => [
+              styles.captureBtn,
+              { backgroundColor: '#3c87f7', opacity: busy ? 0.5 : pressed ? 0.85 : 1 },
+            ]}>
+            <ThemedText style={styles.captureText}>＋ Import</ThemedText>
+          </Pressable>
+        </View>
+
+        {busy && <ActiveJob />}
+        {job.status === 'error' && (
+          <ThemedView type="backgroundElement" style={styles.card}>
+            <ThemedText type="smallBold">Transcription failed</ThemedText>
+            <ThemedText type="small" themeColor="textSecondary">{job.error}</ThemedText>
+          </ThemedView>
+        )}
+
+        {courses.length > 0 && (
+          <ScrollView horizontal showsHorizontalScrollIndicator={false} contentContainerStyle={styles.filterBar}>
+            <FilterChip label="All" active={courseFilter === 'all'} onPress={() => void setCourseFilter('all')} />
+            <FilterChip label="Unsorted" active={courseFilter === null} onPress={() => void setCourseFilter(null)} />
+            {courses.map((c) => (
+              <FilterChip key={c.id} label={c.name} active={courseFilter === c.id} onPress={() => void setCourseFilter(c.id)} />
+            ))}
+          </ScrollView>
+        )}
+
+        <TextInput
+          value={query}
+          onChangeText={(t) => void setQuery(t)}
+          placeholder="Filter by title…"
+          placeholderTextColor={theme.textSecondary}
+          style={[styles.search, { color: theme.text, backgroundColor: theme.backgroundElement }]}
+        />
+
+        {loading && items.length === 0 ? (
+          <ActivityIndicator style={styles.pad} />
+        ) : items.length === 0 ? (
+          <ThemedText type="small" themeColor="textSecondary" style={styles.pad}>
+            {query ? 'No matches.' : 'No lectures yet. Record one or import an audio/video file to begin.'}
+          </ThemedText>
+        ) : (
+          items.map((t) => (
+            <TranscriptRow
+              key={t.id}
+              item={t}
+              courseName={courseName(t.courseId)}
+              onOpen={() => router.push({ pathname: '/transcript/[id]', params: { id: t.id } })}
+              onDelete={() => void remove(t.id)}
+            />
+          ))
+        )}
+      </ScrollView>
+
+      <PreCaptureSheet
+        visible={picked !== null}
+        defaultTitle={picked?.name ?? 'Recording'}
+        onStart={(r) => void onStart(r)}
+        onCancel={() => setPicked(null)}
+      />
+    </ThemedView>
+  );
+}
+
+function FilterChip({ label, active, onPress }: { label: string; active: boolean; onPress: () => void }) {
+  const theme = useTheme();
+  return (
+    <Pressable
+      onPress={onPress}
+      style={[styles.filterChip, { backgroundColor: active ? '#3c87f7' : theme.backgroundElement }]}>
+      <ThemedText type="small" style={active ? styles.chipActive : undefined}>{label}</ThemedText>
+    </Pressable>
+  );
+}
+
+function ActiveJob() {
+  const { stage, progress, partial, chunkIndex, chunkCount, cancel } = useTranscribe();
+  const pct = Math.round(progress * 100);
+  const transcribing = stage === 'transcribing';
+  // While transcribing, show "chunk i/N" so a long file's progress is legible
+  // and it's obvious work is happening between the (discrete) per-chunk updates.
+  const label = stage === 'loading' ? `Loading model… ${pct}%` : `Transcribing… ${pct}%`;
+  const sub =
+    transcribing && chunkCount
+      ? `part ${Math.min((chunkIndex ?? 0) + (pct < 100 ? 1 : 0), chunkCount)} of ${chunkCount}`
+      : undefined;
+  return (
+    <ThemedView type="backgroundElement" style={styles.card}>
+      <View style={styles.rowBetween}>
+        <View style={styles.jobTitleRow}>
+          <ActivityIndicator size="small" />
+          <ThemedText type="smallBold">{label}</ThemedText>
+        </View>
+        <Pressable onPress={cancel} hitSlop={8}>
+          <ThemedText type="small" themeColor="textSecondary">Cancel</ThemedText>
+        </Pressable>
+      </View>
+      <ProgressBar value={progress} />
+      {sub && (
+        <ThemedText type="small" themeColor="textSecondary">{sub}</ThemedText>
+      )}
+      {partial.length > 0 && (
+        <ThemedText type="small" themeColor="textSecondary" numberOfLines={3}>
+          {partial.map((s) => s.text).join(' ')}
+        </ThemedText>
+      )}
+    </ThemedView>
+  );
+}
+
+function ProgressBar({ value }: { value: number }) {
+  return (
+    <View style={styles.track}>
+      <View style={[styles.bar, { width: `${Math.max(2, Math.min(100, value * 100))}%` }]} />
+    </View>
+  );
+}
+
+function TranscriptRow({
+  item,
+  courseName,
+  onOpen,
+  onDelete,
+}: {
+  item: TranscriptMeta;
+  courseName: string | null;
+  onOpen: () => void;
+  onDelete: () => void;
+}) {
+  const date = new Date(item.lectureDate ?? item.createdAt).toLocaleDateString();
+  return (
+    <Pressable onPress={onOpen} style={({ pressed }) => [pressed && styles.pressed]}>
+      <ThemedView type="backgroundElement" style={styles.card}>
+        <View style={styles.rowBetween}>
+          <ThemedText type="smallBold" numberOfLines={1} style={styles.flex}>{item.title}</ThemedText>
+          <Pressable onPress={onDelete} hitSlop={8}>
+            <ThemedText type="small" themeColor="textSecondary">✕</ThemedText>
+          </Pressable>
+        </View>
+        <ThemedText type="small" themeColor="textSecondary">
+          {courseName ? `${courseName} · ` : ''}
+          {date} · {formatClock(item.durationSec)} · {item.segmentCount} segments
+        </ThemedText>
+      </ThemedView>
+    </Pressable>
+  );
+}
+
+const styles = StyleSheet.create({
+  fill: { flex: 1 },
+  content: {
+    padding: Spacing.three,
+    gap: Spacing.three,
+    maxWidth: MaxContentWidth,
+    width: '100%',
+    alignSelf: 'center',
+    paddingBottom: Spacing.six,
+  },
+  flex: { flex: 1 },
+  subHeader: { flexDirection: 'row', alignItems: 'center', gap: Spacing.two },
+  captureRow: { flexDirection: 'row', gap: Spacing.two },
+  captureBtn: { flex: 1, paddingVertical: Spacing.three, borderRadius: Spacing.three, alignItems: 'center' },
+  captureText: { color: '#fff', fontWeight: '700', fontSize: 16 },
+  card: { padding: Spacing.three, borderRadius: Spacing.three, gap: Spacing.two },
+  rowBetween: { flexDirection: 'row', alignItems: 'center', justifyContent: 'space-between', gap: Spacing.two },
+  jobTitleRow: { flexDirection: 'row', alignItems: 'center', gap: Spacing.two, flex: 1 },
+  filterBar: { gap: Spacing.two, paddingVertical: Spacing.one, paddingRight: Spacing.three },
+  filterChip: { paddingHorizontal: Spacing.three, paddingVertical: Spacing.two, borderRadius: 999 },
+  chipActive: { color: '#fff', fontWeight: '700' },
+  search: { borderRadius: Spacing.two, paddingHorizontal: Spacing.three, paddingVertical: Spacing.three, fontSize: 15 },
+  track: { height: 6, borderRadius: 3, backgroundColor: '#88888833', overflow: 'hidden' },
+  bar: { height: 6, borderRadius: 3, backgroundColor: '#3c87f7' },
+  pad: { paddingVertical: Spacing.four, textAlign: 'center' },
+  pressed: { opacity: 0.7 },
+});

src/app/(tabs)/search.tsx

@@ -0,0 +1,231 @@
+import { useFocusEffect, useRouter } from 'expo-router';
+import { useCallback, useEffect, useRef, useState } from 'react';
+import { ActivityIndicator, Pressable, ScrollView, StyleSheet, TextInput, View } from 'react-native';
+
+import { ThemedText } from '@/components/themed-text';
+import { ThemedView } from '@/components/themed-view';
+import { MaxContentWidth, Spacing } from '@/constants/theme';
+import { useTheme } from '@/hooks/use-theme';
+import { formatClock } from '@/lib/format';
+import { searchLectures } from '@/lib/search/search';
+import type { SearchHit } from '@/lib/search/types';
+import { useCourses } from '@/stores/coursesStore';
+import { useEmbedding } from '@/stores/embeddingStore';
+
+const ACCENT = '#3c87f7';
+
+export default function SearchScreen() {
+  const theme = useTheme();
+  const router = useRouter();
+
+  const courses = useCourses((s) => s.items);
+  const refreshCourses = useCourses((s) => s.refresh);
+
+  const { status, progress, pending, refreshPending, buildIndex } = useEmbedding();
+
+  const [query, setQuery] = useState('');
+  const [hits, setHits] = useState<SearchHit[]>([]);
+  const [searching, setSearching] = useState(false);
+  // Distinguish "haven't searched yet" from "searched, got nothing".
+  const [searched, setSearched] = useState(false);
+
+  // Refresh courses + pending count whenever the screen gains focus.
+  useFocusEffect(
+    useCallback(() => {
+      void refreshCourses();
+      void refreshPending();
+    }, [refreshCourses, refreshPending]),
+  );
+
+  // Debounced search on query change. The latest run wins (stale guard via seq).
+  const seq = useRef(0);
+  useEffect(() => {
+    const q = query.trim();
+    if (!q) {
+      setHits([]);
+      setSearched(false);
+      setSearching(false);
+      return;
+    }
+    const mySeq = ++seq.current;
+    setSearching(true);
+    const handle = setTimeout(() => {
+      void searchLectures(q)
+        .then((res) => {
+          if (seq.current !== mySeq) return; // a newer query superseded us
+          setHits(res);
+          setSearched(true);
+        })
+        .catch(() => {
+          if (seq.current !== mySeq) return;
+          setHits([]);
+          setSearched(true);
+        })
+        .finally(() => {
+          if (seq.current !== mySeq) return;
+          setSearching(false);
+        });
+    }, 250);
+    return () => clearTimeout(handle);
+  }, [query]);
+
+  const runNow = useCallback(() => {
+    const q = query.trim();
+    if (!q) return;
+    const mySeq = ++seq.current;
+    setSearching(true);
+    void searchLectures(q)
+      .then((res) => {
+        if (seq.current !== mySeq) return;
+        setHits(res);
+        setSearched(true);
+      })
+      .catch(() => {
+        if (seq.current !== mySeq) return;
+        setHits([]);
+        setSearched(true);
+      })
+      .finally(() => {
+        if (seq.current !== mySeq) return;
+        setSearching(false);
+      });
+  }, [query]);
+
+  const courseName = (cid: string | null) =>
+    cid ? courses.find((c) => c.id === cid)?.name ?? 'Course' : 'Unsorted';
+
+  const openHit = (hit: SearchHit) =>
+    router.push({
+      pathname: '/transcript/[id]',
+      params: { id: hit.transcriptId, t: String(Math.floor(hit.start)) },
+    });
+
+  const indexReady = status === 'ready' && pending === 0;
+
+  return (
+    <ThemedView style={styles.fill}>
+      <ScrollView contentContainerStyle={styles.content} keyboardShouldPersistTaps="handled">
+        <ThemedText type="small" themeColor="textSecondary">
+          Semantic search across your lectures — runs entirely on your device.
+        </ThemedText>
+
+        {pending > 0 && (
+          <ThemedView type="backgroundElement" style={styles.banner}>
+            <ThemedText type="smallBold">
+              Build search index ({pending} {pending === 1 ? 'lecture' : 'lectures'} pending)
+            </ThemedText>
+            {status === 'indexing' ? (
+              <>
+                <ThemedText type="small" themeColor="textSecondary">
+                  Indexing… {Math.round(progress * 100)}%
+                </ThemedText>
+                <ProgressBar value={progress} />
+              </>
+            ) : (
+              <Pressable
+                onPress={() => void buildIndex()}
+                style={({ pressed }) => [styles.bannerBtn, { opacity: pressed ? 0.85 : 1 }]}>
+                <ThemedText style={styles.bannerBtnText}>Build index</ThemedText>
+              </Pressable>
+            )}
+          </ThemedView>
+        )}
+
+        <TextInput
+          value={query}
+          onChangeText={setQuery}
+          onSubmitEditing={runNow}
+          returnKeyType="search"
+          autoFocus
+          placeholder="Search your lectures…"
+          placeholderTextColor={theme.textSecondary}
+          style={[styles.search, { color: theme.text, backgroundColor: theme.backgroundElement }]}
+        />
+
+        {searching ? (
+          <ActivityIndicator style={styles.pad} />
+        ) : query.trim() === '' ? (
+          <ThemedText type="small" themeColor="textSecondary" style={styles.pad}>
+            {indexReady
+              ? 'Type a question or topic to search across your lectures.'
+              : 'Build the index to search.'}
+          </ThemedText>
+        ) : searched && hits.length === 0 ? (
+          <ThemedText type="small" themeColor="textSecondary" style={styles.pad}>
+            No matches.
+          </ThemedText>
+        ) : (
+          hits.map((hit) => (
+            <ResultRow
+              key={`${hit.transcriptId}:${hit.segmentId}`}
+              hit={hit}
+              courseName={courseName(hit.courseId)}
+              onPress={() => openHit(hit)}
+            />
+          ))
+        )}
+      </ScrollView>
+    </ThemedView>
+  );
+}
+
+function ResultRow({
+  hit,
+  courseName,
+  onPress,
+}: {
+  hit: SearchHit;
+  courseName: string;
+  onPress: () => void;
+}) {
+  return (
+    <Pressable onPress={onPress} style={({ pressed }) => [pressed && styles.pressed]}>
+      <ThemedView type="backgroundElement" style={styles.card}>
+        <ThemedText type="small" numberOfLines={2}>
+          {hit.text}
+        </ThemedText>
+        <ThemedText type="small" themeColor="textSecondary">
+          {courseName} · {formatClock(hit.start)}
+        </ThemedText>
+      </ThemedView>
+    </Pressable>
+  );
+}
+
+function ProgressBar({ value }: { value: number }) {
+  return (
+    <View style={styles.track}>
+      <View style={[styles.bar, { width: `${Math.max(2, Math.min(100, value * 100))}%` }]} />
+    </View>
+  );
+}
+
+const styles = StyleSheet.create({
+  fill: { flex: 1 },
+  content: {
+    padding: Spacing.three,
+    gap: Spacing.three,
+    maxWidth: MaxContentWidth,
+    width: '100%',
+    alignSelf: 'center',
+  },
+  banner: { padding: Spacing.three, borderRadius: Spacing.three, gap: Spacing.two },
+  bannerBtn: {
+    backgroundColor: ACCENT,
+    paddingVertical: Spacing.two,
+    borderRadius: Spacing.two,
+    alignItems: 'center',
+  },
+  bannerBtnText: { color: '#fff', fontWeight: '700' },
+  search: {
+    borderRadius: Spacing.two,
+    paddingHorizontal: Spacing.three,
+    paddingVertical: Spacing.two,
+    fontSize: 15,
+  },
+  card: { padding: Spacing.three, borderRadius: Spacing.three, gap: Spacing.two },
+  track: { height: 6, borderRadius: 3, backgroundColor: '#88888833', overflow: 'hidden' },
+  bar: { height: 6, borderRadius: 3, backgroundColor: ACCENT },
+  pad: { paddingVertical: Spacing.four, textAlign: 'center' },
+  pressed: { opacity: 0.7 },
+});

src/app/(tabs)/settings.tsx

@@ -0,0 +1,272 @@
+import { useEffect, useState } from 'react';
+import { ScrollView, StyleSheet, Pressable, TextInput, View } from 'react-native';
+
+import { ThemedText } from '@/components/themed-text';
+import { ThemedView } from '@/components/themed-view';
+import { MaxContentWidth, Spacing } from '@/constants/theme';
+import { useTheme } from '@/hooks/use-theme';
+import { listModels } from '@/lib/models/catalog';
+import { getEngine } from '@/lib/transcription';
+import type { Backend } from '@/lib/types';
+import { useAi } from '@/stores/aiStore';
+import { useTranscribe } from '@/stores/transcribeStore';
+
+const ACCENT = '#3c87f7';
+const DEFAULT_BASE_URL = 'https://api.openai.com/v1';
+const DEFAULT_MODEL = 'gpt-4o-mini';
+
+export default function SettingsScreen() {
+  const theme = useTheme();
+  const modelId = useTranscribe((s) => s.modelId);
+  const setModel = useTranscribe((s) => s.setModel);
+  const models = listModels();
+
+  // Download size depends on the resolved backend: the no-WebGPU WEB path pulls
+  // full-precision (fp32) weights — roughly 2x the quantized figure in the
+  // catalog — while WebGPU (fp16) and native (ggml) are close to it. Resolve the
+  // backend once so the sizes we advertise aren't ~half the real download.
+  const [sizeMult, setSizeMult] = useState(1);
+  useEffect(() => {
+    let alive = true;
+    getEngine()
+      .capabilities()
+      .then((c: { backend: Backend }) => {
+        if (alive) setSizeMult(c.backend === 'wasm' ? 2 : c.backend === 'webgpu' ? 1.4 : 1);
+      })
+      .catch(() => {});
+    return () => {
+      alive = false;
+    };
+  }, []);
+
+  return (
+    <ThemedView style={styles.fill}>
+      <ScrollView contentContainerStyle={styles.content}>
+        <ThemedText type="subtitle">Model</ThemedText>
+        <ThemedText type="small" themeColor="textSecondary">
+          Smaller models are faster and run well on any CPU; larger ones are more accurate but want a GPU.
+          The model downloads once, then works fully offline.
+        </ThemedText>
+
+        {models.map((m) => {
+          const selected = m.id === modelId;
+          return (
+            <Pressable key={m.id} onPress={() => setModel(m.id)}>
+              <ThemedView
+                type={selected ? 'backgroundSelected' : 'backgroundElement'}
+                style={[styles.card, selected && { borderColor: ACCENT, borderWidth: 1 }]}>
+                <View style={styles.rowBetween}>
+                  <ThemedText type="smallBold">{m.label}</ThemedText>
+                  {selected && <ThemedText type="small" style={{ color: ACCENT }}>✓ selected</ThemedText>}
+                </View>
+                <ThemedText type="small" themeColor="textSecondary">
+                  {cap(m.tier)} · ~{Math.round(m.approxMB * sizeMult)} MB · {m.multilingual ? 'multilingual' : 'English-only'}
+                </ThemedText>
+              </ThemedView>
+            </Pressable>
+          );
+        })}
+
+        <View style={styles.spacer} />
+        <AiSection />
+
+        <View style={styles.spacer} />
+        <ThemedText type="subtitle">Privacy</ThemedText>
+        <ThemedText type="small" themeColor="textSecondary">
+          Wisp transcribes entirely on your device using OpenAI&apos;s Whisper model. Your audio is never
+          uploaded to any server — there is no account, no per-minute fee, and it keeps working with the
+          network off. The only download is the model file itself.
+        </ThemedText>
+      </ScrollView>
+    </ThemedView>
+  );
+}
+
+function AiSection() {
+  const theme = useTheme();
+
+  const cloud = useAi((s) => s.cloud);
+  const engineKind = useAi((s) => s.engineKind);
+  const setCloud = useAi((s) => s.setCloud);
+  const refreshAvailability = useAi((s) => s.refreshAvailability);
+
+  // Form fields seeded from any saved cloud config.
+  const [baseUrl, setBaseUrl] = useState(cloud?.baseUrl ?? DEFAULT_BASE_URL);
+  const [model, setModel] = useState(cloud?.model ?? DEFAULT_MODEL);
+  const [apiKey, setApiKey] = useState(cloud?.apiKey ?? '');
+
+  // Probe availability (WebGPU / key set) when the screen mounts.
+  useEffect(() => {
+    void refreshAvailability();
+  }, [refreshAvailability]);
+
+  // Keep the form in sync if the saved config changes elsewhere.
+  useEffect(() => {
+    setBaseUrl(cloud?.baseUrl ?? DEFAULT_BASE_URL);
+    setModel(cloud?.model ?? DEFAULT_MODEL);
+    setApiKey(cloud?.apiKey ?? '');
+  }, [cloud]);
+
+  const engineLabel =
+    engineKind === 'cloud'
+      ? 'Cloud (your key)'
+      : engineKind === 'webllm'
+        ? 'On-device (WebGPU)'
+        : 'Not available';
+
+  const onSave = () => {
+    const url = baseUrl.trim() || DEFAULT_BASE_URL;
+    const mdl = model.trim() || DEFAULT_MODEL;
+    const key = apiKey.trim();
+    if (!key) return;
+    setCloud({ baseUrl: url, apiKey: key, model: mdl });
+    void refreshAvailability();
+  };
+
+  const onClear = () => {
+    setCloud(undefined);
+    setApiKey('');
+    setBaseUrl(DEFAULT_BASE_URL);
+    setModel(DEFAULT_MODEL);
+    void refreshAvailability();
+  };
+
+  return (
+    <>
+      <ThemedText type="subtitle">AI (optional)</ThemedText>
+      <ThemedText type="small" themeColor="textSecondary">
+        Wisp can answer questions about your lectures with cited answers. This is fully optional —
+        without it, you still get on-device semantic search. It runs on-device when your browser
+        supports WebGPU, or through your own API key below.
+      </ThemedText>
+
+      <ThemedView type="backgroundElement" style={styles.card}>
+        <View style={styles.rowBetween}>
+          <ThemedText type="smallBold">Active engine</ThemedText>
+          <ThemedText type="small" style={{ color: engineKind === 'none' ? theme.textSecondary : ACCENT }}>
+            {engineLabel}
+          </ThemedText>
+        </View>
+        <ThemedText type="small" themeColor="textSecondary">
+          On-device generation needs a WebGPU-capable browser. If that isn&apos;t available, add an
+          API key below to use a cloud model instead.
+        </ThemedText>
+      </ThemedView>
+
+      <ThemedText type="smallBold">Bring your own key</ThemedText>
+      <ThemedText type="small" themeColor="textSecondary">
+        Use any OpenAI-compatible endpoint. Your key is stored locally on this device and is never
+        shared with Wisp. Only your question and the retrieved lecture snippets are sent to the
+        endpoint — never your raw audio or full transcripts.
+      </ThemedText>
+
+      <Field label="Base URL">
+        <TextInput
+          value={baseUrl}
+          onChangeText={setBaseUrl}
+          autoCapitalize="none"
+          autoCorrect={false}
+          keyboardType="url"
+          placeholder={DEFAULT_BASE_URL}
+          placeholderTextColor={theme.textSecondary}
+          style={[styles.input, { color: theme.text, backgroundColor: theme.backgroundElement }]}
+        />
+      </Field>
+
+      <Field label="Model">
+        <TextInput
+          value={model}
+          onChangeText={setModel}
+          autoCapitalize="none"
+          autoCorrect={false}
+          placeholder={DEFAULT_MODEL}
+          placeholderTextColor={theme.textSecondary}
+          style={[styles.input, { color: theme.text, backgroundColor: theme.backgroundElement }]}
+        />
+      </Field>
+
+      <Field label="API key">
+        <TextInput
+          value={apiKey}
+          onChangeText={setApiKey}
+          autoCapitalize="none"
+          autoCorrect={false}
+          secureTextEntry
+          placeholder="sk-…"
+          placeholderTextColor={theme.textSecondary}
+          style={[styles.input, { color: theme.text, backgroundColor: theme.backgroundElement }]}
+        />
+      </Field>
+
+      <View style={styles.btnRow}>
+        <Pressable
+          onPress={onSave}
+          disabled={apiKey.trim() === ''}
+          style={({ pressed }) => [
+            styles.saveBtn,
+            { opacity: apiKey.trim() === '' ? 0.5 : pressed ? 0.85 : 1 },
+          ]}>
+          <ThemedText style={styles.saveBtnText}>Save key</ThemedText>
+        </Pressable>
+        <Pressable
+          onPress={onClear}
+          disabled={!cloud}
+          style={({ pressed }) => [
+            styles.clearBtn,
+            { borderColor: theme.textSecondary, opacity: !cloud ? 0.5 : pressed ? 0.85 : 1 },
+          ]}>
+          <ThemedText type="smallBold" themeColor="textSecondary">
+            Clear
+          </ThemedText>
+        </Pressable>
+      </View>
+    </>
+  );
+}
+
+function Field({ label, children }: { label: string; children: React.ReactNode }) {
+  return (
+    <View style={styles.field}>
+      <ThemedText type="small" themeColor="textSecondary">
+        {label}
+      </ThemedText>
+      {children}
+    </View>
+  );
+}
+
+function cap(s: string) {
+  return s.charAt(0).toUpperCase() + s.slice(1);
+}
+
+const styles = StyleSheet.create({
+  fill: { flex: 1 },
+  content: { padding: Spacing.three, gap: Spacing.two, maxWidth: MaxContentWidth, width: '100%', alignSelf: 'center' },
+  card: { padding: Spacing.three, borderRadius: Spacing.three, gap: Spacing.one },
+  rowBetween: { flexDirection: 'row', alignItems: 'center', justifyContent: 'space-between' },
+  spacer: { height: Spacing.three },
+  field: { gap: Spacing.one },
+  input: {
+    borderRadius: Spacing.two,
+    paddingHorizontal: Spacing.three,
+    paddingVertical: Spacing.two,
+    fontSize: 15,
+  },
+  btnRow: { flexDirection: 'row', gap: Spacing.two, marginTop: Spacing.one },
+  saveBtn: {
+    flex: 1,
+    backgroundColor: ACCENT,
+    paddingVertical: Spacing.two,
+    borderRadius: Spacing.two,
+    alignItems: 'center',
+  },
+  saveBtnText: { color: '#fff', fontWeight: '700' },
+  clearBtn: {
+    paddingHorizontal: Spacing.four,
+    paddingVertical: Spacing.two,
+    borderRadius: Spacing.two,
+    borderWidth: 1,
+    alignItems: 'center',
+    justifyContent: 'center',
+  },
+});

src/app/(tabs)/study.tsx

@@ -0,0 +1,200 @@
+import { useFocusEffect, useRouter } from 'expo-router';
+import { useCallback, useState } from 'react';
+import { ActivityIndicator, Pressable, ScrollView, StyleSheet, View } from 'react-native';
+
+import { ThemedText } from '@/components/themed-text';
+import { ThemedView } from '@/components/themed-view';
+import { MaxContentWidth, Spacing } from '@/constants/theme';
+import { useTheme } from '@/hooks/use-theme';
+import { getRepo, type Flashcard } from '@/lib/db';
+import { downloadText } from '@/lib/download';
+import { review } from '@/lib/learn';
+import { useCourses } from '@/stores/coursesStore';
+
+const ACCENT = '#3c87f7';
+
+// Course scope for the due queue: 'all' = no filter, null = Unsorted, string = course id.
+type Scope = 'all' | string | null;
+
+const GRADES: { label: string; grade: 0 | 1 | 2 | 3 }[] = [
+  { label: 'Again', grade: 0 },
+  { label: 'Hard', grade: 1 },
+  { label: 'Good', grade: 2 },
+  { label: 'Easy', grade: 3 },
+];
+
+export default function StudyScreen() {
+  const theme = useTheme();
+  const router = useRouter();
+
+  const courses = useCourses((s) => s.items);
+  const refreshCourses = useCourses((s) => s.refresh);
+
+  const [scope, setScope] = useState<Scope>('all');
+  const [queue, setQueue] = useState<Flashcard[] | null>(null);
+  const [index, setIndex] = useState(0);
+  const [revealed, setRevealed] = useState(false);
+
+  const load = useCallback(async (s: Scope) => {
+    setQueue(null);
+    setIndex(0);
+    setRevealed(false);
+    const cards = await getRepo().listDueFlashcards(
+      s === 'all' ? {} : { courseId: s },
+    );
+    setQueue(cards);
+  }, []);
+
+  // Reload the due queue and courses every time the screen gains focus.
+  useFocusEffect(
+    useCallback(() => {
+      void refreshCourses();
+      void load(scope);
+    }, [refreshCourses, load, scope]),
+  );
+
+  const pickScope = (s: Scope) => {
+    setScope(s);
+    void load(s);
+  };
+
+  const total = queue?.length ?? 0;
+  const card = queue && index < total ? queue[index] : undefined;
+
+  const grade = async (g: 0 | 1 | 2 | 3) => {
+    if (!card) return;
+    await getRepo().updateFlashcardSrs(card.id, review(card.srs, g, Date.now()));
+    setRevealed(false);
+    setIndex((i) => i + 1);
+  };
+
+  const exportAnki = async () => {
+    const all = await getRepo().listFlashcards();
+    const rows = all.map((c) => `${csvField(c.front)},${csvField(c.back)}`);
+    const csv = ['front,back', ...rows].join('\r\n');
+    downloadText('wisp-flashcards.csv', 'text/csv', csv);
+  };
+
+  const courseName = (cid?: string | null) =>
+    cid ? courses.find((c) => c.id === cid)?.name ?? 'Course' : 'Unsorted';
+
+  return (
+    <ThemedView style={styles.fill}>
+      <ScrollView contentContainerStyle={styles.content}>
+        <View style={styles.topActions}>
+          <Pressable onPress={() => void exportAnki()} hitSlop={8}>
+            <ThemedText type="link" themeColor="textSecondary">Export Anki (.csv)</ThemedText>
+          </Pressable>
+        </View>
+        <ScrollView horizontal showsHorizontalScrollIndicator={false} contentContainerStyle={styles.filterBar}>
+          <FilterChip label="All" active={scope === 'all'} onPress={() => pickScope('all')} />
+          <FilterChip label="Unsorted" active={scope === null} onPress={() => pickScope(null)} />
+          {courses.map((c) => (
+            <FilterChip key={c.id} label={c.name} active={scope === c.id} onPress={() => pickScope(c.id)} />
+          ))}
+        </ScrollView>
+
+        {queue === null ? (
+          <ActivityIndicator style={styles.pad} />
+        ) : card === undefined ? (
+          <ThemedText type="small" themeColor="textSecondary" style={styles.pad}>
+            {total === 0
+              ? 'No cards due — generate some from a lecture.'
+              : 'All done — no more cards due right now.'}
+          </ThemedText>
+        ) : (
+          <>
+            <ThemedText type="small" themeColor="textSecondary">
+              {index + 1} of {total}
+            </ThemedText>
+
+            <ThemedView type="backgroundElement" style={styles.card}>
+              <ThemedText type="small" themeColor="textSecondary">{courseName(card.courseId)}</ThemedText>
+              <ThemedText type="subtitle" style={styles.front}>{card.front}</ThemedText>
+
+              {revealed ? (
+                <>
+                  <View style={[styles.divider, { backgroundColor: theme.backgroundSelected }]} />
+                  <ThemedText type="default">{card.back}</ThemedText>
+                  {card.start !== undefined && (
+                    <Pressable
+                      onPress={() =>
+                        router.push({
+                          pathname: '/transcript/[id]',
+                          params: { id: card.transcriptId, t: String(Math.floor(card.start ?? 0)) },
+                        })
+                      }
+                      hitSlop={6}>
+                      <ThemedText type="linkPrimary">Jump to lecture</ThemedText>
+                    </Pressable>
+                  )}
+                </>
+              ) : (
+                <Pressable
+                  onPress={() => setRevealed(true)}
+                  style={({ pressed }) => [styles.showBtn, { backgroundColor: ACCENT, opacity: pressed ? 0.85 : 1 }]}>
+                  <ThemedText style={styles.showBtnText}>Show answer</ThemedText>
+                </Pressable>
+              )}
+            </ThemedView>
+
+            {revealed && (
+              <View style={styles.gradeRow}>
+                {GRADES.map((g) => (
+                  <Pressable
+                    key={g.grade}
+                    onPress={() => void grade(g.grade)}
+                    style={({ pressed }) => [
+                      styles.gradeBtn,
+                      { backgroundColor: theme.backgroundElement, opacity: pressed ? 0.7 : 1 },
+                    ]}>
+                    <ThemedText type="smallBold">{g.label}</ThemedText>
+                  </Pressable>
+                ))}
+              </View>
+            )}
+          </>
+        )}
+      </ScrollView>
+    </ThemedView>
+  );
+}
+
+function FilterChip({ label, active, onPress }: { label: string; active: boolean; onPress: () => void }) {
+  const theme = useTheme();
+  return (
+    <Pressable
+      onPress={onPress}
+      style={[styles.filterChip, { backgroundColor: active ? ACCENT : theme.backgroundElement }]}>
+      <ThemedText type="small" style={active ? styles.chipActive : undefined}>{label}</ThemedText>
+    </Pressable>
+  );
+}
+
+/** RFC4180-quote a CSV field: wrap in quotes, double any embedded quotes. */
+function csvField(value: string): string {
+  return `"${value.replace(/"/g, '""')}"`;
+}
+
+const styles = StyleSheet.create({
+  fill: { flex: 1 },
+  content: {
+    padding: Spacing.three,
+    gap: Spacing.three,
+    maxWidth: MaxContentWidth,
+    width: '100%',
+    alignSelf: 'center',
+  },
+  filterBar: { gap: Spacing.two, paddingVertical: Spacing.one, paddingRight: Spacing.three },
+  filterChip: { paddingHorizontal: Spacing.three, paddingVertical: Spacing.one, borderRadius: 999 },
+  topActions: { flexDirection: 'row', justifyContent: 'flex-end' },
+  chipActive: { color: '#fff', fontWeight: '700' },
+  card: { padding: Spacing.four, borderRadius: Spacing.three, gap: Spacing.three },
+  front: { fontSize: 24, lineHeight: 32 },
+  divider: { height: StyleSheet.hairlineWidth, marginVertical: Spacing.one },
+  showBtn: { paddingVertical: Spacing.three, borderRadius: Spacing.three, alignItems: 'center' },
+  showBtnText: { color: '#fff', fontWeight: '700', fontSize: 16 },
+  gradeRow: { flexDirection: 'row', gap: Spacing.two },
+  gradeBtn: { flex: 1, paddingVertical: Spacing.three, borderRadius: Spacing.two, alignItems: 'center' },
+  pad: { paddingVertical: Spacing.four, textAlign: 'center' },
+});

src/app/_layout.tsx

@@ -14,9 +14,12 @@ export default function RootLayout() {
   return (
     <ThemeProvider value={scheme === 'dark' ? DarkTheme : DefaultTheme}>
       <Stack>
-        <Stack.Screen name="index" options={{ title: 'Wisp' }} />
+        <Stack.Screen name="(tabs)" options={{ headerShown: false }} />
+        <Stack.Screen name="record" options={{ title: 'Record' }} />
         <Stack.Screen name="transcript/[id]" options={{ title: 'Transcript' }} />
-        <Stack.Screen name="settings" options={{ title: 'Settings' }} />
+        <Stack.Screen name="courses" options={{ title: 'Courses' }} />
+        <Stack.Screen name="quiz" options={{ title: 'Quiz' }} />
+        <Stack.Screen name="bibliography" options={{ title: 'Bibliography' }} />
       </Stack>
       <StatusBar style="auto" />
     </ThemeProvider>

src/app/bibliography.tsx

@@ -0,0 +1,220 @@
+import * as Linking from 'expo-linking';
+import { Stack, useLocalSearchParams } from 'expo-router';
+import { useCallback, useState } from 'react';
+import { ActivityIndicator, Pressable, ScrollView, StyleSheet, View } from 'react-native';
+
+import { ThemedText } from '@/components/themed-text';
+import { ThemedView } from '@/components/themed-view';
+import { MaxContentWidth, Spacing } from '@/constants/theme';
+import { useTheme } from '@/hooks/use-theme';
+import { getRepo } from '@/lib/db';
+import { downloadText } from '@/lib/download';
+import {
+  detectCitations,
+  lookupLinksFor,
+  resolveCitation,
+  toBibtex,
+  toRis,
+  type Citation,
+  type LookupLinks,
+  type RefMetadata,
+} from '@/lib/enrich';
+
+const ACCENT = '#3c87f7';
+
+/** A detected citation paired with its (best-effort) resolved metadata. */
+interface BibEntry {
+  citation: Citation;
+  meta?: RefMetadata;
+}
+
+export default function BibliographyScreen() {
+  const theme = useTheme();
+  const { courseId } = useLocalSearchParams<{ courseId?: string }>();
+
+  const [entries, setEntries] = useState<BibEntry[] | null>(null);
+  const [building, setBuilding] = useState(false);
+
+  // Collect citations across the course's (or all) transcripts, then resolve
+  // each via the metadata APIs. Detection is local; lookups send only ids.
+  const build = useCallback(async () => {
+    setBuilding(true);
+    setEntries(null);
+    try {
+      const repo = getRepo();
+      const metas = courseId ? await repo.listByCourse(courseId) : await repo.list();
+
+      // Detect citations across every transcript body. Dedupe by kind+value so
+      // a reference cited in several lectures shows once.
+      const byKey = new Map<string, Citation>();
+      for (const m of metas) {
+        const full = await repo.get(m.id);
+        if (!full) continue;
+        for (const c of detectCitations(full.segments)) {
+          const key = `${c.kind}:${c.value}`;
+          if (!byKey.has(key)) byKey.set(key, c);
+        }
+      }
+      const citations = [...byKey.values()];
+      // Seed the list immediately (links work without lookups), then resolve.
+      setEntries(citations.map((citation) => ({ citation })));
+
+      const resolved = await Promise.all(
+        citations.map(async (citation) => {
+          try {
+            return { citation, meta: await resolveCitation(citation) };
+          } catch {
+            return { citation };
+          }
+        }),
+      );
+      setEntries(resolved);
+    } finally {
+      setBuilding(false);
+    }
+  }, [courseId]);
+
+  // Resolved metadata feeds the exporters; fall back to the raw citation text
+  // for entries that couldn't be looked up so nothing is silently dropped.
+  const exportable = (entries ?? [])
+    .map((e) => e.meta ?? fallbackMeta(e.citation))
+    .filter((m): m is RefMetadata => m !== undefined);
+
+  const exportBibtex = () => {
+    if (exportable.length === 0) return;
+    downloadText('bibliography.bib', 'application/x-bibtex', toBibtex(exportable));
+  };
+
+  const exportRis = () => {
+    if (exportable.length === 0) return;
+    downloadText('bibliography.ris', 'application/x-research-info-systems', toRis(exportable));
+  };
+
+  const openLink = (url: string) => {
+    void Linking.openURL(url);
+  };
+
+  return (
+    <ThemedView style={styles.fill}>
+      <Stack.Screen options={{ title: 'Bibliography' }} />
+      <ScrollView contentContainerStyle={styles.content}>
+        <ThemedText type="small" themeColor="textSecondary">
+          {courseId
+            ? 'References detected across this course’s lectures.'
+            : 'References detected across all lectures.'}
+        </ThemedText>
+
+        <View style={styles.row}>
+          <Pressable
+            onPress={() => void build()}
+            disabled={building}
+            style={({ pressed }) => [
+              styles.btn,
+              { backgroundColor: ACCENT, opacity: building ? 0.6 : pressed ? 0.85 : 1 },
+            ]}>
+            <ThemedText style={styles.btnText}>
+              {building ? 'Scanning…' : entries ? 'Re-scan' : 'Build bibliography'}
+            </ThemedText>
+          </Pressable>
+          {entries && entries.length > 0 && (
+            <>
+              <Pressable
+                onPress={exportBibtex}
+                style={({ pressed }) => [
+                  styles.btn,
+                  { backgroundColor: theme.backgroundElement, opacity: pressed ? 0.7 : 1 },
+                ]}>
+                <ThemedText type="smallBold">Export BibTeX</ThemedText>
+              </Pressable>
+              <Pressable
+                onPress={exportRis}
+                style={({ pressed }) => [
+                  styles.btn,
+                  { backgroundColor: theme.backgroundElement, opacity: pressed ? 0.7 : 1 },
+                ]}>
+                <ThemedText type="smallBold">Export RIS</ThemedText>
+              </Pressable>
+            </>
+          )}
+        </View>
+
+        {building && entries === null && <ActivityIndicator style={styles.pad} />}
+
+        {entries && entries.length === 0 && (
+          <ThemedText type="small" themeColor="textSecondary" style={styles.pad}>
+            No references found.
+          </ThemedText>
+        )}
+
+        {entries?.map((e, i) => {
+          const links = lookupLinksFor(e.citation, e.meta);
+          return (
+            <ThemedView key={`${e.citation.kind}:${e.citation.value}:${i}`} type="backgroundElement" style={styles.card}>
+              {e.meta?.title ? (
+                <ThemedText type="smallBold">{e.meta.title}</ThemedText>
+              ) : (
+                <ThemedText type="smallBold">{e.citation.raw}</ThemedText>
+              )}
+              {(e.meta?.authors?.length || e.meta?.year) && (
+                <ThemedText type="small" themeColor="textSecondary">
+                  {[e.meta?.authors?.join(', '), e.meta?.year].filter(Boolean).join(' · ')}
+                </ThemedText>
+              )}
+              <ThemedText type="small" themeColor="textSecondary">
+                {e.meta?.source ?? e.citation.kind}
+              </ThemedText>
+              <View style={styles.linkChips}>
+                {linkEntries(links).map(([label, url]) => (
+                  <Pressable
+                    key={label}
+                    onPress={() => openLink(url)}
+                    style={({ pressed }) => [
+                      styles.linkChip,
+                      { backgroundColor: theme.backgroundSelected, opacity: pressed ? 0.7 : 1 },
+                    ]}>
+                    <ThemedText type="small">{label}</ThemedText>
+                  </Pressable>
+                ))}
+              </View>
+            </ThemedView>
+          );
+        })}
+      </ScrollView>
+    </ThemedView>
+  );
+}
+
+/** Minimal metadata so an unresolved citation still exports something useful. */
+function fallbackMeta(c: Citation): RefMetadata {
+  return { title: c.raw, source: c.kind };
+}
+
+/** Ordered, labeled, defined link entries from a LookupLinks bag. */
+function linkEntries(links: LookupLinks): [string, string][] {
+  const order: [keyof LookupLinks, string][] = [
+    ['doi', 'DOI'],
+    ['arxiv', 'arXiv'],
+    ['openAlex', 'OpenAlex'],
+    ['openLibrary', 'Open Library'],
+    ['googleBooks', 'Google Books'],
+    ['wikipedia', 'Wikipedia'],
+    ['annasArchive', "Anna's Archive"],
+    ['libgen', 'LibGen'],
+  ];
+  return order
+    .map(([key, label]) => [label, links[key]] as const)
+    .filter((pair): pair is [string, string] => typeof pair[1] === 'string' && pair[1].length > 0)
+    .map(([label, url]) => [label, url]);
+}
+
+const styles = StyleSheet.create({
+  fill: { flex: 1 },
+  content: { padding: Spacing.three, gap: Spacing.two, maxWidth: MaxContentWidth, width: '100%', alignSelf: 'center' },
+  row: { flexDirection: 'row', flexWrap: 'wrap', gap: Spacing.two, alignItems: 'center' },
+  btn: { paddingHorizontal: Spacing.three, paddingVertical: Spacing.two, borderRadius: Spacing.two, alignItems: 'center' },
+  btnText: { color: '#fff', fontWeight: '700', fontSize: 14 },
+  card: { padding: Spacing.three, borderRadius: Spacing.three, gap: 2 },
+  linkChips: { flexDirection: 'row', flexWrap: 'wrap', gap: Spacing.one, marginTop: Spacing.one },
+  linkChip: { paddingHorizontal: Spacing.two, paddingVertical: 2, borderRadius: 999 },
+  pad: { paddingVertical: Spacing.four, textAlign: 'center' },
+});

src/app/courses.tsx

@@ -0,0 +1,159 @@
+import { Link, Stack, useFocusEffect } from 'expo-router';
+import { useCallback, useState } from 'react';
+import { Pressable, ScrollView, StyleSheet, TextInput, View } from 'react-native';
+
+import { ThemedText } from '@/components/themed-text';
+import { ThemedView } from '@/components/themed-view';
+import { MaxContentWidth, Spacing } from '@/constants/theme';
+import { useTheme } from '@/hooks/use-theme';
+import { getRepo } from '@/lib/db';
+import { useCourses } from '@/stores/coursesStore';
+
+/** Per-course rollup shown on each row. */
+interface CourseStat {
+  count: number;
+  hours: number;
+}
+
+export default function CoursesScreen() {
+  const theme = useTheme();
+  const { items, refresh, createCourse, rename, remove } = useCourses();
+  const [name, setName] = useState('');
+  const [editing, setEditing] = useState<string | null>(null);
+  const [editName, setEditName] = useState('');
+  const [stats, setStats] = useState<Record<string, CourseStat>>({});
+
+  // Load the courses list, then roll up lecture count + total hours per course
+  // in a single pass (one listByCourse() per course; fetched once on focus).
+  const reload = useCallback(async () => {
+    await refresh();
+    const courses = useCourses.getState().items;
+    const entries = await Promise.all(
+      courses.map(async (c) => {
+        const metas = await getRepo().listByCourse(c.id);
+        const hours = metas.reduce((sum, m) => sum + m.durationSec, 0) / 3600;
+        return [c.id, { count: metas.length, hours }] as const;
+      }),
+    );
+    setStats(Object.fromEntries(entries));
+  }, [refresh]);
+
+  useFocusEffect(
+    useCallback(() => {
+      void reload();
+    }, [reload]),
+  );
+
+  const add = async () => {
+    const n = name.trim();
+    if (!n) return;
+    await createCourse({ name: n });
+    setName('');
+    await reload();
+  };
+
+  return (
+    <ThemedView style={styles.fill}>
+      <Stack.Screen options={{ title: 'Courses' }} />
+      <ScrollView contentContainerStyle={styles.content} keyboardShouldPersistTaps="handled">
+        <ThemedText type="small" themeColor="textSecondary">
+          Group lectures by course. Deleting a course moves its transcripts to Unsorted (nothing is lost).
+        </ThemedText>
+
+        <View style={styles.row}>
+          <TextInput
+            value={name}
+            onChangeText={setName}
+            placeholder="New course name"
+            placeholderTextColor={theme.textSecondary}
+            style={[styles.input, styles.flex, { color: theme.text, backgroundColor: theme.backgroundElement }]}
+          />
+          <Pressable onPress={() => void add()} style={[styles.addBtn, { opacity: name.trim() ? 1 : 0.5 }]}>
+            <ThemedText style={styles.white}>Add</ThemedText>
+          </Pressable>
+        </View>
+
+        {items.length === 0 && (
+          <ThemedText type="small" themeColor="textSecondary" style={styles.pad}>No courses yet.</ThemedText>
+        )}
+
+        {items.map((c) => (
+          <ThemedView key={c.id} type="backgroundElement" style={styles.card}>
+            {editing === c.id ? (
+              <View style={styles.row}>
+                <TextInput
+                  value={editName}
+                  onChangeText={setEditName}
+                  autoFocus
+                  style={[styles.input, styles.flex, { color: theme.text, backgroundColor: theme.background }]}
+                />
+                <Pressable
+                  onPress={async () => {
+                    if (editName.trim()) await rename(c.id, editName.trim());
+                    setEditing(null);
+                  }}
+                  style={styles.addBtn}>
+                  <ThemedText style={styles.white}>Save</ThemedText>
+                </Pressable>
+              </View>
+            ) : (
+              <View style={styles.rowBetween}>
+                <View style={styles.flex}>
+                  <ThemedText type="smallBold" numberOfLines={1}>{c.name}</ThemedText>
+                  <ThemedText type="small" themeColor="textSecondary">
+                    {statLine(stats[c.id])}
+                  </ThemedText>
+                </View>
+                <View style={styles.actions}>
+                  <Link href={{ pathname: '/bibliography', params: { courseId: c.id } }} asChild>
+                    <Pressable hitSlop={8}>
+                      <ThemedText type="small" themeColor="textSecondary">Bibliography</ThemedText>
+                    </Pressable>
+                  </Link>
+                  <Pressable
+                    onPress={() => {
+                      setEditing(c.id);
+                      setEditName(c.name);
+                    }}
+                    hitSlop={8}>
+                    <ThemedText type="small" themeColor="textSecondary">Rename</ThemedText>
+                  </Pressable>
+                  <Pressable
+                    onPress={async () => {
+                      await remove(c.id);
+                      await reload();
+                    }}
+                    hitSlop={8}>
+                    <ThemedText type="small" themeColor="textSecondary">Delete</ThemedText>
+                  </Pressable>
+                </View>
+              </View>
+            )}
+          </ThemedView>
+        ))}
+      </ScrollView>
+    </ThemedView>
+  );
+}
+
+/** "3 lectures · 4.2 h" — falls back to a zero state while stats load. */
+function statLine(stat: CourseStat | undefined): string {
+  const count = stat?.count ?? 0;
+  const hours = stat?.hours ?? 0;
+  const lectures = `${count} ${count === 1 ? 'lecture' : 'lectures'}`;
+  return `${lectures} · ${hours.toFixed(1)} h`;
+}
+
+const styles = StyleSheet.create({
+  fill: { flex: 1 },
+  content: { padding: Spacing.three, gap: Spacing.two, maxWidth: MaxContentWidth, width: '100%', alignSelf: 'center' },
+  row: { flexDirection: 'row', gap: Spacing.two, alignItems: 'center' },
+  rowBetween: { flexDirection: 'row', alignItems: 'center', justifyContent: 'space-between', gap: Spacing.two },
+  flex: { flex: 1 },
+  input: { borderRadius: Spacing.two, paddingHorizontal: Spacing.three, paddingVertical: Spacing.two, fontSize: 15 },
+  addBtn: { backgroundColor: '#3c87f7', paddingHorizontal: Spacing.three, paddingVertical: Spacing.two, borderRadius: Spacing.two },
+  white: { color: '#fff', fontWeight: '700' },
+  card: { padding: Spacing.three, borderRadius: Spacing.three },
+  actions: { flexDirection: 'row', gap: Spacing.three },
+  pad: { paddingVertical: Spacing.four, textAlign: 'center' },
+});

src/app/index.tsx

@@ -1,182 +0,0 @@
-import { Link, useFocusEffect, useRouter } from 'expo-router';
-import { useCallback } from 'react';
-import {
-  ActivityIndicator,
-  Platform,
-  Pressable,
-  ScrollView,
-  StyleSheet,
-  TextInput,
-  View,
-} from 'react-native';
-
-import { ThemedText } from '@/components/themed-text';
-import { ThemedView } from '@/components/themed-view';
-import { MaxContentWidth, Spacing } from '@/constants/theme';
-import { useTheme } from '@/hooks/use-theme';
-import { formatClock } from '@/lib/format';
-import { MODELS } from '@/lib/models/catalog';
-import { pickAudio } from '@/lib/pickAudio';
-import type { TranscriptMeta } from '@/lib/db';
-import { useTranscribe } from '@/stores/transcribeStore';
-import { useTranscripts } from '@/stores/transcriptsStore';
-
-export default function LibraryScreen() {
-  const theme = useTheme();
-  const router = useRouter();
-  const { items, loading, query, setQuery, refresh, remove } = useTranscripts();
-  const job = useTranscribe();
-
-  // Refresh the list whenever the screen regains focus (e.g. after a transcribe).
-  useFocusEffect(
-    useCallback(() => {
-      void refresh();
-    }, [refresh]),
-  );
-
-  const busy = job.status === 'loading' || job.status === 'transcribing';
-
-  const onNew = useCallback(async () => {
-    if (busy) return;
-    const picked = await pickAudio();
-    if (!picked) return;
-    const id = await useTranscribe.getState().start(picked, { title: picked.name });
-    if (id) router.push({ pathname: '/transcript/[id]', params: { id } });
-  }, [busy, router]);
-
-  return (
-    <ThemedView style={styles.fill}>
-      <ScrollView contentContainerStyle={styles.content}>
-        <View style={styles.headerRow}>
-          <View style={styles.flex}>
-            <ThemedText type="subtitle">Wisp</ThemedText>
-            <ThemedText type="small" themeColor="textSecondary">
-              Private transcription — runs on your device, nothing uploaded.
-            </ThemedText>
-          </View>
-          <Link href="/settings" asChild>
-            <Pressable hitSlop={10}>
-              <ThemedText type="link" themeColor="textSecondary">
-                ⚙ Settings
-              </ThemedText>
-            </Pressable>
-          </Link>
-        </View>
-
-        <Pressable
-          onPress={onNew}
-          disabled={busy}
-          style={({ pressed }) => [
-            styles.newButton,
-            { backgroundColor: '#3c87f7', opacity: busy ? 0.5 : pressed ? 0.85 : 1 },
-          ]}>
-          <ThemedText style={styles.newButtonText}>＋ New transcription</ThemedText>
-        </Pressable>
-
-        {busy && <ActiveJob />}
-        {job.status === 'error' && (
-          <ThemedView type="backgroundElement" style={styles.card}>
-            <ThemedText type="smallBold">Transcription failed</ThemedText>
-            <ThemedText type="small" themeColor="textSecondary">
-              {job.error}
-            </ThemedText>
-          </ThemedView>
-        )}
-
-        <TextInput
-          value={query}
-          onChangeText={(t) => void setQuery(t)}
-          placeholder="Search transcripts…"
-          placeholderTextColor={theme.textSecondary}
-          style={[styles.search, { color: theme.text, backgroundColor: theme.backgroundElement }]}
-        />
-
-        {loading && items.length === 0 ? (
-          <ActivityIndicator style={styles.pad} />
-        ) : items.length === 0 ? (
-          <ThemedText type="small" themeColor="textSecondary" style={styles.pad}>
-            {query ? 'No matches.' : 'No transcripts yet. Pick an audio or video file to begin.'}
-          </ThemedText>
-        ) : (
-          items.map((t) => (
-            <TranscriptRow key={t.id} item={t} onOpen={() => router.push({ pathname: '/transcript/[id]', params: { id: t.id } })} onDelete={() => void remove(t.id)} />
-          ))
-        )}
-      </ScrollView>
-    </ThemedView>
-  );
-}
-
-function ActiveJob() {
-  const { stage, progress, partial, cancel } = useTranscribe();
-  const pct = Math.round(progress * 100);
-  return (
-    <ThemedView type="backgroundElement" style={styles.card}>
-      <View style={styles.rowBetween}>
-        <ThemedText type="smallBold">
-          {stage === 'loading' ? 'Loading model…' : 'Transcribing…'} {pct}%
-        </ThemedText>
-        <Pressable onPress={cancel} hitSlop={8}>
-          <ThemedText type="small" themeColor="textSecondary">Cancel</ThemedText>
-        </Pressable>
-      </View>
-      <ProgressBar value={progress} />
-      {partial.length > 0 && (
-        <ThemedText type="small" themeColor="textSecondary" numberOfLines={3}>
-          {partial.map((s) => s.text).join(' ')}
-        </ThemedText>
-      )}
-    </ThemedView>
-  );
-}
-
-function ProgressBar({ value }: { value: number }) {
-  return (
-    <View style={styles.track}>
-      <View style={[styles.bar, { width: `${Math.max(2, Math.min(100, value * 100))}%` }]} />
-    </View>
-  );
-}
-
-function TranscriptRow({ item, onOpen, onDelete }: { item: TranscriptMeta; onOpen: () => void; onDelete: () => void }) {
-  const date = new Date(item.createdAt).toLocaleDateString();
-  return (
-    <Pressable onPress={onOpen} style={({ pressed }) => [pressed && styles.pressed]}>
-      <ThemedView type="backgroundElement" style={styles.card}>
-        <View style={styles.rowBetween}>
-          <ThemedText type="smallBold" numberOfLines={1} style={styles.flex}>
-            {item.title}
-          </ThemedText>
-          <Pressable onPress={onDelete} hitSlop={8}>
-            <ThemedText type="small" themeColor="textSecondary">✕</ThemedText>
-          </Pressable>
-        </View>
-        <ThemedText type="small" themeColor="textSecondary">
-          {date} · {formatClock(item.durationSec)} · {MODELS[item.modelId]?.label ?? item.modelId} · {item.segmentCount} segments
-        </ThemedText>
-      </ThemedView>
-    </Pressable>
-  );
-}
-
-const styles = StyleSheet.create({
-  fill: { flex: 1 },
-  content: {
-    padding: Spacing.three,
-    gap: Spacing.three,
-    maxWidth: MaxContentWidth,
-    width: '100%',
-    alignSelf: 'center',
-  },
-  flex: { flex: 1 },
-  headerRow: { flexDirection: 'row', alignItems: 'flex-start', gap: Spacing.two },
-  newButton: { paddingVertical: Spacing.three, borderRadius: Spacing.three, alignItems: 'center' },
-  newButtonText: { color: '#fff', fontWeight: '700', fontSize: 16 },
-  card: { padding: Spacing.three, borderRadius: Spacing.three, gap: Spacing.two },
-  rowBetween: { flexDirection: 'row', alignItems: 'center', justifyContent: 'space-between', gap: Spacing.two },
-  search: { borderRadius: Spacing.two, paddingHorizontal: Spacing.three, paddingVertical: Spacing.two, fontSize: 15 },
-  track: { height: 6, borderRadius: 3, backgroundColor: '#88888833', overflow: 'hidden' },
-  bar: { height: 6, borderRadius: 3, backgroundColor: '#3c87f7' },
-  pad: { paddingVertical: Spacing.four, textAlign: 'center' },
-  pressed: { opacity: 0.7 },
-});

src/app/quiz.tsx

@@ -0,0 +1,188 @@
+import { Stack, useLocalSearchParams, useRouter } from 'expo-router';
+import { useEffect, useMemo, useState } from 'react';
+import { ActivityIndicator, Pressable, ScrollView, StyleSheet, View } from 'react-native';
+
+import { ThemedText } from '@/components/themed-text';
+import { ThemedView } from '@/components/themed-view';
+import { MaxContentWidth, Spacing } from '@/constants/theme';
+import { useTheme } from '@/hooks/use-theme';
+import { getRepo, type Transcript } from '@/lib/db';
+import { formatClock } from '@/lib/format';
+import { generateQuiz, glossary, type QuizQuestion } from '@/lib/learn';
+
+const ACCENT = '#3c87f7';
+const CORRECT = '#30a46c';
+const WRONG = '#e5484d';
+
+export default function QuizScreen() {
+  const theme = useTheme();
+  const router = useRouter();
+  const { id } = useLocalSearchParams<{ id: string }>();
+
+  const [transcript, setTranscript] = useState<Transcript | null | undefined>(undefined);
+
+  useEffect(() => {
+    let alive = true;
+    void getRepo()
+      .get(id)
+      .then((t) => {
+        if (alive) setTranscript(t ?? null);
+      });
+    return () => {
+      alive = false;
+    };
+  }, [id]);
+
+  const questions = useMemo<QuizQuestion[]>(
+    () => (transcript ? generateQuiz(glossary(transcript.segments)) : []),
+    [transcript],
+  );
+
+  const [index, setIndex] = useState(0);
+  const [picked, setPicked] = useState<number | null>(null);
+  const [score, setScore] = useState(0);
+
+  const total = questions.length;
+  const q = index < total ? questions[index] : undefined;
+  const finished = total > 0 && index >= total;
+
+  const onPick = (optionIndex: number) => {
+    if (picked !== null || !q) return;
+    setPicked(optionIndex);
+    if (optionIndex === q.answerIndex) setScore((s) => s + 1);
+  };
+
+  const next = () => {
+    setPicked(null);
+    setIndex((i) => i + 1);
+  };
+
+  const restart = () => {
+    setIndex(0);
+    setPicked(null);
+    setScore(0);
+  };
+
+  if (transcript === undefined) {
+    return (
+      <Centered>
+        <ActivityIndicator />
+      </Centered>
+    );
+  }
+
+  return (
+    <ThemedView style={styles.fill}>
+      <Stack.Screen options={{ title: 'Quiz' }} />
+      <ScrollView contentContainerStyle={styles.content}>
+        {transcript === null ? (
+          <ThemedText type="small" themeColor="textSecondary" style={styles.pad}>
+            Transcript not found.
+          </ThemedText>
+        ) : total === 0 ? (
+          <ThemedText type="small" themeColor="textSecondary" style={styles.pad}>
+            Not enough glossary terms to build a quiz for this lecture.
+          </ThemedText>
+        ) : finished ? (
+          <ThemedView type="backgroundElement" style={styles.card}>
+            <ThemedText type="subtitle">Quiz complete</ThemedText>
+            <ThemedText type="default">
+              You scored {score} of {total}.
+            </ThemedText>
+            <Pressable
+              onPress={restart}
+              style={({ pressed }) => [styles.primaryBtn, { backgroundColor: ACCENT, opacity: pressed ? 0.85 : 1 }]}>
+              <ThemedText style={styles.primaryBtnText}>Try again</ThemedText>
+            </Pressable>
+            <Pressable
+              onPress={() => router.push({ pathname: '/transcript/[id]', params: { id } })}
+              style={({ pressed }) => [styles.secondaryBtn, { backgroundColor: theme.backgroundSelected, opacity: pressed ? 0.85 : 1 }]}>
+              <ThemedText type="smallBold">Back to lecture</ThemedText>
+            </Pressable>
+          </ThemedView>
+        ) : q ? (
+          <>
+            <ThemedText type="small" themeColor="textSecondary">
+              Question {index + 1} of {total}
+            </ThemedText>
+            <ThemedText type="subtitle" style={styles.question}>{q.question}</ThemedText>
+
+            <View style={styles.options}>
+              {q.options.map((opt, i) => {
+                const isAnswer = i === q.answerIndex;
+                const isPicked = i === picked;
+                const revealed = picked !== null;
+                const border =
+                  revealed && isAnswer ? CORRECT : revealed && isPicked ? WRONG : theme.backgroundElement;
+                return (
+                  <Pressable
+                    key={i}
+                    onPress={() => onPick(i)}
+                    disabled={revealed}
+                    style={({ pressed }) => [
+                      styles.option,
+                      { backgroundColor: theme.backgroundElement, borderColor: border, opacity: pressed && !revealed ? 0.8 : 1 },
+                    ]}>
+                    <ThemedText type="small">{opt}</ThemedText>
+                  </Pressable>
+                );
+              })}
+            </View>
+
+            {picked !== null && (
+              <ThemedView type="backgroundElement" style={styles.feedback}>
+                <ThemedText type="smallBold" themeColor="textSecondary">
+                  {picked === q.answerIndex ? 'Correct' : 'Incorrect'}
+                </ThemedText>
+                {q.start !== undefined && (
+                  <Pressable
+                    onPress={() =>
+                      router.push({
+                        pathname: '/transcript/[id]',
+                        params: { id, t: String(Math.floor(q.start ?? 0)) },
+                      })
+                    }
+                    hitSlop={6}>
+                    <ThemedText type="linkPrimary">Jump to lecture ({formatClock(q.start)})</ThemedText>
+                  </Pressable>
+                )}
+                <Pressable
+                  onPress={next}
+                  style={({ pressed }) => [styles.primaryBtn, { backgroundColor: ACCENT, opacity: pressed ? 0.85 : 1 }]}>
+                  <ThemedText style={styles.primaryBtnText}>
+                    {index + 1 >= total ? 'See score' : 'Next'}
+                  </ThemedText>
+                </Pressable>
+              </ThemedView>
+            )}
+          </>
+        ) : null}
+      </ScrollView>
+    </ThemedView>
+  );
+}
+
+function Centered({ children }: { children: React.ReactNode }) {
+  return <ThemedView style={[styles.fill, styles.centered]}>{children}</ThemedView>;
+}
+
+const styles = StyleSheet.create({
+  fill: { flex: 1 },
+  centered: { alignItems: 'center', justifyContent: 'center' },
+  content: {
+    padding: Spacing.three,
+    gap: Spacing.three,
+    maxWidth: MaxContentWidth,
+    width: '100%',
+    alignSelf: 'center',
+  },
+  card: { padding: Spacing.four, borderRadius: Spacing.three, gap: Spacing.three },
+  question: { fontSize: 24, lineHeight: 32 },
+  options: { gap: Spacing.two },
+  option: { padding: Spacing.three, borderRadius: Spacing.three, borderWidth: 2 },
+  feedback: { padding: Spacing.three, borderRadius: Spacing.three, gap: Spacing.two },
+  primaryBtn: { paddingVertical: Spacing.three, borderRadius: Spacing.three, alignItems: 'center' },
+  primaryBtnText: { color: '#fff', fontWeight: '700', fontSize: 16 },
+  secondaryBtn: { paddingVertical: Spacing.three, borderRadius: Spacing.three, alignItems: 'center' },
+  pad: { paddingVertical: Spacing.four, textAlign: 'center' },
+});

src/app/record.tsx

@@ -0,0 +1,143 @@
+import { Stack, useRouter } from 'expo-router';
+import { useState } from 'react';
+import { Pressable, StyleSheet, View } from 'react-native';
+
+import { PreCaptureSheet, type PreCaptureResult } from '@/components/PreCaptureSheet';
+import { ThemedText } from '@/components/themed-text';
+import { ThemedView } from '@/components/themed-view';
+import { MaxContentWidth, Spacing } from '@/constants/theme';
+import { useTheme } from '@/hooks/use-theme';
+import { useRecorder, type RecordingResult } from '@/hooks/useRecorder';
+import { formatClock } from '@/lib/format';
+import { useTranscribe } from '@/stores/transcribeStore';
+
+export default function RecordScreen() {
+  const theme = useTheme();
+  const router = useRouter();
+  const rec = useRecorder();
+  const [pending, setPending] = useState<RecordingResult | null>(null);
+
+  const onStop = async () => {
+    const r = await rec.stop();
+    if (r) setPending(r);
+    else router.back();
+  };
+
+  const onStart = async (meta: PreCaptureResult) => {
+    const r = pending;
+    setPending(null);
+    if (!r) return;
+    const id = await useTranscribe.getState().start({ data: r.data, name: r.name }, meta);
+    if (id) router.replace({ pathname: '/transcript/[id]', params: { id } });
+    else router.back();
+  };
+
+  if (!rec.supported) {
+    return (
+      <ThemedView style={[styles.fill, styles.center]}>
+        <Stack.Screen options={{ title: 'Record' }} />
+        <ThemedText type="small" themeColor="textSecondary" style={styles.notice}>
+          In-app recording is available in the browser for now. On mobile, use your phone&apos;s
+          recorder and import the file (native live capture is coming).
+        </ThemedText>
+      </ThemedView>
+    );
+  }
+
+  const recording = rec.status === 'recording';
+  const paused = rec.status === 'paused';
+
+  return (
+    <ThemedView style={[styles.fill, styles.center]}>
+      <Stack.Screen options={{ title: 'Record' }} />
+      <View style={styles.stage}>
+        <ThemedText style={styles.timer}>{formatClock(Math.floor(rec.elapsedMs / 1000))}</ThemedText>
+        <ThemedText type="small" themeColor="textSecondary">
+          {recording ? '● Recording — on your device' : paused ? 'Paused' : 'Ready to record'}
+        </ThemedText>
+
+        <View style={[styles.meterTrack, { backgroundColor: theme.backgroundElement }]}>
+          <View
+            style={[
+              styles.meterFill,
+              { width: `${Math.round((recording ? rec.level : 0) * 100)}%`, backgroundColor: '#e5484d' },
+            ]}
+          />
+        </View>
+
+        {rec.error && (
+          <ThemedText type="small" style={styles.error}>{rec.error}</ThemedText>
+        )}
+
+        <View style={styles.controls}>
+          {rec.status === 'idle' || rec.status === 'stopped' ? (
+            <BigButton label="Start recording" color="#e5484d" onPress={() => void rec.start()} />
+          ) : (
+            <>
+              {recording ? (
+                <BigButton label="Pause" color={theme.backgroundElement} textColor={theme.text} onPress={rec.pause} />
+              ) : (
+                <BigButton label="Resume" color="#3c87f7" onPress={rec.resume} />
+              )}
+              <BigButton label="Stop & transcribe" color="#3c87f7" onPress={() => void onStop()} />
+            </>
+          )}
+        </View>
+
+        <ThemedText type="small" themeColor="textSecondary" style={styles.hint}>
+          Keep this tab in front while recording. Everything stays on your device.
+        </ThemedText>
+      </View>
+
+      <PreCaptureSheet
+        visible={pending !== null}
+        defaultTitle={defaultName()}
+        onStart={(m) => void onStart(m)}
+        onCancel={() => {
+          setPending(null);
+          router.back();
+        }}
+      />
+    </ThemedView>
+  );
+}
+
+function BigButton({
+  label,
+  color,
+  textColor,
+  onPress,
+}: {
+  label: string;
+  color: string;
+  textColor?: string;
+  onPress: () => void;
+}) {
+  return (
+    <Pressable onPress={onPress} style={({ pressed }) => [styles.btn, { backgroundColor: color, opacity: pressed ? 0.85 : 1 }]}>
+      <ThemedText style={[styles.btnText, textColor ? { color: textColor } : styles.btnTextLight]}>{label}</ThemedText>
+    </Pressable>
+  );
+}
+
+function defaultName(): string {
+  const d = new Date();
+  const pad = (n: number) => String(n).padStart(2, '0');
+  return `Lecture ${d.getFullYear()}-${pad(d.getMonth() + 1)}-${pad(d.getDate())}`;
+}
+
+const styles = StyleSheet.create({
+  fill: { flex: 1 },
+  center: { alignItems: 'center', justifyContent: 'center' },
+  stage: { width: '100%', maxWidth: MaxContentWidth, padding: Spacing.four, gap: Spacing.three, alignItems: 'center' },
+  timer: { fontSize: 56, fontWeight: '700', fontVariant: ['tabular-nums'] },
+  meterTrack: { width: '80%', height: 8, borderRadius: 4, overflow: 'hidden' },
+  meterFill: { height: 8, borderRadius: 4 },
+  controls: { flexDirection: 'row', gap: Spacing.two, marginTop: Spacing.three, flexWrap: 'wrap', justifyContent: 'center' },
+  btn: { paddingVertical: Spacing.three, paddingHorizontal: Spacing.four, borderRadius: Spacing.three, alignItems: 'center' },
+  btnText: { fontWeight: '700', fontSize: 16 },
+  btnTextLight: { color: '#fff' },
+  error: { color: '#e5484d' },
+  notice: { padding: Spacing.four, textAlign: 'center', maxWidth: 480 },
+  hint: { textAlign: 'center', marginTop: Spacing.two },
+});

src/app/settings.tsx

@@ -1,66 +0,0 @@
-import { ScrollView, StyleSheet, Pressable, View } from 'react-native';
-
-import { ThemedText } from '@/components/themed-text';
-import { ThemedView } from '@/components/themed-view';
-import { MaxContentWidth, Spacing } from '@/constants/theme';
-import { useTheme } from '@/hooks/use-theme';
-import { listModels } from '@/lib/models/catalog';
-import { useTranscribe } from '@/stores/transcribeStore';
-
-export default function SettingsScreen() {
-  const theme = useTheme();
-  const modelId = useTranscribe((s) => s.modelId);
-  const setModel = useTranscribe((s) => s.setModel);
-  const models = listModels();
-
-  return (
-    <ThemedView style={styles.fill}>
-      <ScrollView contentContainerStyle={styles.content}>
-        <ThemedText type="subtitle">Model</ThemedText>
-        <ThemedText type="small" themeColor="textSecondary">
-          Smaller models are faster and run well on any CPU; larger ones are more accurate but want a GPU.
-          The model downloads once, then works fully offline.
-        </ThemedText>
-
-        {models.map((m) => {
-          const selected = m.id === modelId;
-          return (
-            <Pressable key={m.id} onPress={() => setModel(m.id)}>
-              <ThemedView
-                type={selected ? 'backgroundSelected' : 'backgroundElement'}
-                style={[styles.card, selected && { borderColor: '#3c87f7', borderWidth: 1 }]}>
-                <View style={styles.rowBetween}>
-                  <ThemedText type="smallBold">{m.label}</ThemedText>
-                  {selected && <ThemedText type="small" style={{ color: '#3c87f7' }}>✓ selected</ThemedText>}
-                </View>
-                <ThemedText type="small" themeColor="textSecondary">
-                  {cap(m.tier)} · ~{m.approxMB} MB · {m.multilingual ? 'multilingual' : 'English-only'}
-                </ThemedText>
-              </ThemedView>
-            </Pressable>
-          );
-        })}
-
-        <View style={styles.spacer} />
-        <ThemedText type="subtitle">Privacy</ThemedText>
-        <ThemedText type="small" themeColor="textSecondary">
-          Wisp transcribes entirely on your device using OpenAI&apos;s Whisper model. Your audio is never
-          uploaded to any server — there is no account, no per-minute fee, and it keeps working with the
-          network off. The only download is the model file itself.
-        </ThemedText>
-      </ScrollView>
-    </ThemedView>
-  );
-}
-
-function cap(s: string) {
-  return s.charAt(0).toUpperCase() + s.slice(1);
-}
-
-const styles = StyleSheet.create({
-  fill: { flex: 1 },
-  content: { padding: Spacing.three, gap: Spacing.two, maxWidth: MaxContentWidth, width: '100%', alignSelf: 'center' },
-  card: { padding: Spacing.three, borderRadius: Spacing.three, gap: Spacing.one },
-  rowBetween: { flexDirection: 'row', alignItems: 'center', justifyContent: 'space-between' },
-  spacer: { height: Spacing.three },
-});

src/app/transcript/[id].tsx

@@ -1,4 +1,5 @@
-import { Stack, useLocalSearchParams } from 'expo-router';
+import * as Linking from 'expo-linking';
+import { Link, Stack, useLocalSearchParams } from 'expo-router';
 import { useEffect, useMemo, useRef, useState } from 'react';
 import {
   ActivityIndicator,
@@ -16,26 +17,100 @@ import { MaxContentWidth, Spacing } from '@/constants/theme';
 import { useTheme } from '@/hooks/use-theme';
 import { getRepo, type Transcript } from '@/lib/db';
 import { downloadText } from '@/lib/download';
+import { errorMessage } from '@/lib/errorMessage';
+import {
+  detectCitations,
+  detectEvents,
+  lookupLinksFor,
+  resolveCitation,
+  toIcs,
+  type CandidateEvent,
+  type Citation,
+  type LookupLinks,
+  type RefMetadata,
+} from '@/lib/enrich';
 import { EXPORT_META, formatTranscript, type ExportFormat } from '@/lib/export';
 import { formatClock } from '@/lib/format';
+import {
+  cardsFromGlossary,
+  glossary,
+  summarize,
+  type GlossaryEntry,
+  type SourcedSentence,
+} from '@/lib/learn';
 import type { Segment } from '@/lib/types';
 import { useTranscribe } from '@/stores/transcribeStore';
 
+const ACCENT = '#3c87f7';
+
 export default function TranscriptScreen() {
   const theme = useTheme();
-  const { id } = useLocalSearchParams<{ id: string }>();
+  const { id, t } = useLocalSearchParams<{ id: string; t?: string }>();
+  // Deep-link target time (seconds) from a search hit, if any.
+  const jumpTo = useMemo(() => {
+    const n = Number(t);
+    return Number.isFinite(n) && n >= 0 ? n : null;
+  }, [t]);
 
   const [transcript, setTranscript] = useState<Transcript | null | undefined>(undefined);
+  const [loadError, setLoadError] = useState<string | null>(null);
   const [title, setTitle] = useState('');
   const [segments, setSegments] = useState<Segment[]>([]);
   const [dirty, setDirty] = useState(false);
   const [saving, setSaving] = useState(false);
   const [currentTime, setCurrentTime] = useState(0);
 
-  // The just-transcribed audio is playable in-session (object URL on web).
+  // Study aids (deterministic, computed on demand — no storage, no model).
-  const audioUrl = useTranscribe((s) => (s.lastTranscriptId === id ? s.audioUrl : undefined));
+  const [aids, setAids] = useState<{
+    summary: SourcedSentence[];
+    glossary: GlossaryEntry[];
+  } | null>(null);
+  const [creatingCards, setCreatingCards] = useState(false);
+  const [cardsAdded, setCardsAdded] = useState<number | null>(null);
+
+  // Dates & references (Phase 4) — detected on demand, nothing persisted.
+  // `enrichment === null` means "not scanned yet".
+  const [enrichment, setEnrichment] = useState<{
+    events: CandidateEvent[];
+    citations: Citation[];
+  } | null>(null);
+  // Per-citation resolved metadata + in-flight flags, keyed by list index.
+  const [refMeta, setRefMeta] = useState<Record<number, RefMetadata>>({});
+  const [refLoading, setRefLoading] = useState<Record<number, boolean>>({});
+
+  const scrollRef = useRef<ScrollView | null>(null);
+  // Y offset of each rendered segment row, captured via onLayout, for scroll-to.
+  const segmentYs = useRef<Record<number, number>>({});
+  // Ensure we only auto-jump once per (id, t) deep-link.
+  const jumpedRef = useRef(false);
+
+  // Prefer the just-transcribed in-session audio; otherwise load the persisted
+  // source media so playback/scrub works after a reload (ROADMAP Phase 0).
+  const sessionAudioUrl = useTranscribe((s) => (s.lastTranscriptId === id ? s.audioUrl : undefined));
+  const [persistedUrl, setPersistedUrl] = useState<string | undefined>(undefined);
+  const audioUrl = sessionAudioUrl ?? persistedUrl;
   const audioRef = useRef<HTMLAudioElement | null>(null);
 
+  useEffect(() => {
+    if (sessionAudioUrl) return; // already have in-session audio
+    let url: string | undefined;
+    let alive = true;
+    void getRepo()
+      .getMediaUrl(id)
+      .then((u) => {
+        if (!alive) return;
+        url = u;
+        setPersistedUrl(u);
+      })
+      .catch(() => {
+        // Missing/failed media is non-fatal — playback just stays unavailable.
+      });
+    return () => {
+      alive = false;
+      if (url && Platform.OS === 'web' && url.startsWith('blob:')) URL.revokeObjectURL(url);
+    };
+  }, [id, sessionAudioUrl]);
+
   useEffect(() => {
     let alive = true;
     void getRepo()
@@ -45,6 +120,12 @@ export default function TranscriptScreen() {
         setTranscript(t ?? null);
         setTitle(t?.title ?? '');
         setSegments(t?.segments ?? []);
+      })
+      .catch((e) => {
+        // Without this, a rejected load left the screen on its spinner forever.
+        if (!alive) return;
+        setLoadError(errorMessage(e));
+        setTranscript(null);
       });
     return () => {
       alive = false;
@@ -70,13 +151,48 @@ export default function TranscriptScreen() {
     [segments, currentTime],
   );
 
-  const seek = (t: number) => {
+  const seek = (time: number) => {
     const el = audioRef.current;
     if (!el) return;
-    el.currentTime = t;
+    el.currentTime = time;
     void el.play();
   };
 
+  // Reset the one-shot jump guard whenever the deep-link target changes.
+  useEffect(() => {
+    jumpedRef.current = false;
+  }, [id, jumpTo]);
+
+  // Deep-link jump: once the transcript is loaded and we have a finite `t`,
+  // seek the audio (if present), highlight the matching segment, and scroll it
+  // into view. Works even without audio (scroll/highlight by time only).
+  useEffect(() => {
+    if (jumpTo === null || jumpedRef.current) return;
+    if (!transcript || segments.length === 0) return;
+
+    const target = segments.findIndex((s) => jumpTo >= s.start && jumpTo < s.end);
+    const idx = target >= 0 ? target : nearestIndex(segments, jumpTo);
+    if (idx < 0) return;
+
+    jumpedRef.current = true;
+    // Drive the highlight by time even if audio isn't available.
+    setCurrentTime(jumpTo);
+    if (audioRef.current) seek(jumpTo);
+
+    // Scroll once the row's Y offset has been measured (next frames).
+    const scrollToIdx = () => {
+      const y = segmentYs.current[idx];
+      if (y === undefined) return false;
+      scrollRef.current?.scrollTo({ y: Math.max(0, y - Spacing.four), animated: true });
+      return true;
+    };
+    if (!scrollToIdx()) {
+      // Layout may not be measured yet; retry on the next ticks.
+      const timers = [requestAnimationFrame(() => { if (!scrollToIdx()) setTimeout(scrollToIdx, 120); })];
+      return () => timers.forEach(cancelAnimationFrame);
+    }
+  }, [jumpTo, transcript, segments, audioUrl]);
+
   const editSegment = (i: number, text: string) => {
     setSegments((prev) => prev.map((s, idx) => (idx === i ? { ...s, text } : s)));
     setDirty(true);
@@ -99,22 +215,93 @@ export default function TranscriptScreen() {
     downloadText(`${safeName}.${meta.ext}`, meta.mime, content);
   };
 
+  // Compute summary + glossary on demand (pure helpers, nothing persisted).
+  const generateAids = () => {
+    setAids({ summary: summarize(segments), glossary: glossary(segments) });
+    setCardsAdded(null);
+  };
+
+  // Seeds derived from the current glossary — drives the "Create N flashcards" label.
+  const cardSeeds = useMemo(
+    () => (aids ? cardsFromGlossary(aids.glossary) : []),
+    [aids],
+  );
+
+  const createCards = async () => {
+    if (cardSeeds.length === 0 || creatingCards) return;
+    setCreatingCards(true);
+    try {
+      const created = await getRepo().createFlashcards(
+        cardSeeds.map((s) => ({
+          transcriptId: id,
+          courseId: transcript?.courseId ?? null,
+          segmentId: s.segmentId,
+          start: s.start,
+          front: s.front,
+          back: s.back,
+        })),
+      );
+      setCardsAdded(created.length);
+    } finally {
+      setCreatingCards(false);
+    }
+  };
+
+  // Scan segments for deadlines + citations. Dates are anchored to the lecture
+  // date when known, else the transcript's creation time. Pure & local.
+  const scanEnrichment = () => {
+    const anchorMs = transcript?.lectureDate ?? transcript?.createdAt ?? Date.now();
+    setEnrichment({
+      events: detectEvents(segments, anchorMs),
+      citations: detectCitations(segments),
+    });
+    setRefMeta({});
+    setRefLoading({});
+  };
+
+  // Download a calendar file for one or more events.
+  const exportIcs = (events: CandidateEvent[]) => {
+    if (events.length === 0) return;
+    const ics = toIcs(
+      events.map((e) => ({ title: e.title, dateMs: e.dateMs, allDay: e.allDay })),
+    );
+    downloadText('event.ics', 'text/calendar', ics);
+  };
+
+  // Network lookup for a single citation (sends only the id string, never text).
+  const lookupRef = async (i: number, c: Citation) => {
+    if (refLoading[i]) return;
+    setRefLoading((prev) => ({ ...prev, [i]: true }));
+    try {
+      const meta = await resolveCitation(c);
+      if (meta) setRefMeta((prev) => ({ ...prev, [i]: meta }));
+    } finally {
+      setRefLoading((prev) => ({ ...prev, [i]: false }));
+    }
+  };
+
+  const openLink = (url: string) => {
+    void Linking.openURL(url);
+  };
+
   if (transcript === undefined) return <Centered><ActivityIndicator /></Centered>;
   if (transcript === null)
     return (
       <Centered>
-        <ThemedText type="small" themeColor="textSecondary">Transcript not found.</ThemedText>
+        <ThemedText type="small" themeColor="textSecondary">
+          {loadError ? `Couldn't load this transcript: ${loadError}` : 'Transcript not found.'}
+        </ThemedText>
       </Centered>
     );
 
   return (
     <ThemedView style={styles.fill}>
       <Stack.Screen options={{ title: title || 'Transcript' }} />
-      <ScrollView contentContainerStyle={styles.content}>
+      <ScrollView ref={scrollRef} contentContainerStyle={styles.content}>
         <TextInput
           value={title}
-          onChangeText={(t) => {
+          onChangeText={(next) => {
-            setTitle(t);
+            setTitle(next);
             setDirty(true);
           }}
           style={[styles.title, { color: theme.text }]}
@@ -124,7 +311,7 @@ export default function TranscriptScreen() {
 
         {!audioUrl && (
           <ThemedText type="small" themeColor="textSecondary">
-            Tip: audio playback is available right after transcribing. Re-import the file to scrub along.
+            No source audio stored for this transcript — playback unavailable.
           </ThemedText>
         )}
 
@@ -139,8 +326,226 @@ export default function TranscriptScreen() {
           ))}
         </View>
 
+        {/* Study aids — deterministic helpers, computed on demand. */}
+        <ThemedView type="backgroundElement" style={styles.aidsCard}>
+          <ThemedText type="smallBold">Study aids</ThemedText>
+          <View style={styles.aidsRow}>
+            <Pressable
+              onPress={generateAids}
+              disabled={segments.length === 0}
+              style={({ pressed }) => [
+                styles.aidBtn,
+                { backgroundColor: ACCENT, opacity: segments.length === 0 ? 0.5 : pressed ? 0.85 : 1 },
+              ]}>
+              <ThemedText style={styles.aidBtnText}>
+                {aids ? 'Regenerate summary & glossary' : 'Generate summary & glossary'}
+              </ThemedText>
+            </Pressable>
+            <Link href={{ pathname: '/quiz', params: { id } }} asChild>
+              <Pressable
+                style={({ pressed }) => [
+                  styles.aidBtn,
+                  { backgroundColor: theme.backgroundSelected, opacity: pressed ? 0.85 : 1 },
+                ]}>
+                <ThemedText style={styles.aidBtnTextAlt}>Quiz this lecture</ThemedText>
+              </Pressable>
+            </Link>
+          </View>
+
+          {aids && (
+            <>
+              {aids.summary.length > 0 ? (
+                <View style={styles.aidSection}>
+                  <ThemedText type="smallBold" themeColor="textSecondary">Summary</ThemedText>
+                  {aids.summary.map((s, i) => (
+                    <Pressable key={`sum-${i}`} onPress={() => seek(s.start)} hitSlop={4}>
+                      <View style={styles.aidLine}>
+                        <ThemedText type="code" themeColor="textSecondary" style={styles.ts}>
+                          {formatClock(s.start)}
+                        </ThemedText>
+                        <ThemedText type="small" style={styles.flex}>{s.text}</ThemedText>
+                      </View>
+                    </Pressable>
+                  ))}
+                </View>
+              ) : (
+                <ThemedText type="small" themeColor="textSecondary">No summary could be derived.</ThemedText>
+              )}
+
+              {aids.glossary.length > 0 ? (
+                <View style={styles.aidSection}>
+                  <ThemedText type="smallBold" themeColor="textSecondary">Glossary</ThemedText>
+                  {aids.glossary.map((g, i) => (
+                    <Pressable key={`glo-${i}`} onPress={() => seek(g.start)} hitSlop={4}>
+                      <View style={styles.aidLine}>
+                        <ThemedText type="code" themeColor="textSecondary" style={styles.ts}>
+                          {formatClock(g.start)}
+                        </ThemedText>
+                        <ThemedText type="small" style={styles.flex}>
+                          <ThemedText type="smallBold">{g.term}</ThemedText>
+                          {` — ${g.definition}`}
+                        </ThemedText>
+                      </View>
+                    </Pressable>
+                  ))}
+                </View>
+              ) : (
+                <ThemedText type="small" themeColor="textSecondary">No glossary terms found.</ThemedText>
+              )}
+
+              {cardSeeds.length > 0 && (
+                <Pressable
+                  onPress={() => void createCards()}
+                  disabled={creatingCards}
+                  style={({ pressed }) => [
+                    styles.aidBtn,
+                    { backgroundColor: ACCENT, opacity: creatingCards ? 0.6 : pressed ? 0.85 : 1 },
+                  ]}>
+                  <ThemedText style={styles.aidBtnText}>
+                    {creatingCards ? 'Adding…' : `Create ${cardSeeds.length} flashcards`}
+                  </ThemedText>
+                </Pressable>
+              )}
+
+              {cardsAdded !== null && (
+                <ThemedText type="small" themeColor="textSecondary">
+                  Added {cardsAdded} {cardsAdded === 1 ? 'card' : 'cards'}.
+                </ThemedText>
+              )}
+            </>
+          )}
+        </ThemedView>
+
+        {/* Dates & references — detected on demand (pure/local detection). */}
+        <ThemedView type="backgroundElement" style={styles.aidsCard}>
+          <ThemedText type="smallBold">Dates & references</ThemedText>
+          <Pressable
+            onPress={scanEnrichment}
+            disabled={segments.length === 0}
+            style={({ pressed }) => [
+              styles.aidBtn,
+              { backgroundColor: ACCENT, opacity: segments.length === 0 ? 0.5 : pressed ? 0.85 : 1 },
+            ]}>
+            <ThemedText style={styles.aidBtnText}>
+              {enrichment ? 'Re-scan for dates & references' : 'Scan for dates & references'}
+            </ThemedText>
+          </Pressable>
+
+          {enrichment && (
+            <>
+              {/* --- Dates --- */}
+              <View style={styles.aidSection}>
+                <View style={styles.refHeader}>
+                  <ThemedText type="smallBold" themeColor="textSecondary">Dates</ThemedText>
+                  {enrichment.events.length > 0 && (
+                    <Pressable onPress={() => exportIcs(enrichment.events)} hitSlop={6}>
+                      <ThemedText type="linkPrimary">Add all ({enrichment.events.length})</ThemedText>
+                    </Pressable>
+                  )}
+                </View>
+                {enrichment.events.length === 0 ? (
+                  <ThemedText type="small" themeColor="textSecondary">No dates found.</ThemedText>
+                ) : (
+                  enrichment.events.map((e, i) => (
+                    <View key={`evt-${i}`} style={styles.refRow}>
+                      <Pressable
+                        onPress={() => e.start !== undefined && seek(e.start)}
+                        disabled={e.start === undefined}
+                        hitSlop={4}
+                        style={styles.flex}>
+                        <ThemedText type="small">{e.title}</ThemedText>
+                        <ThemedText type="small" themeColor="textSecondary">
+                          {formatEventDate(e)}
+                          {e.start !== undefined ? ` · ${formatClock(e.start)}` : ''}
+                        </ThemedText>
+                      </Pressable>
+                      <Pressable onPress={() => exportIcs([e])} hitSlop={6}>
+                        <ThemedText type="link" themeColor="textSecondary">Add to calendar</ThemedText>
+                      </Pressable>
+                    </View>
+                  ))
+                )}
+              </View>
+
+              {/* --- References --- */}
+              <View style={styles.aidSection}>
+                <ThemedText type="smallBold" themeColor="textSecondary">References</ThemedText>
+                {enrichment.citations.length === 0 ? (
+                  <ThemedText type="small" themeColor="textSecondary">No references found.</ThemedText>
+                ) : (
+                  enrichment.citations.map((c, i) => {
+                    const meta = refMeta[i];
+                    const links = lookupLinksFor(c, meta);
+                    return (
+                      <View key={`cit-${i}`} style={styles.citCard}>
+                        <View style={styles.refRow}>
+                          <Pressable
+                            onPress={() => c.start !== undefined && seek(c.start)}
+                            disabled={c.start === undefined}
+                            hitSlop={4}
+                            style={styles.flex}>
+                            <ThemedText type="small">{c.raw}</ThemedText>
+                            {c.start !== undefined && (
+                              <ThemedText type="code" themeColor="textSecondary">
+                                {formatClock(c.start)}
+                              </ThemedText>
+                            )}
+                          </Pressable>
+                          <Pressable
+                            onPress={() => void lookupRef(i, c)}
+                            disabled={refLoading[i]}
+                            hitSlop={6}>
+                            {refLoading[i] ? (
+                              <ActivityIndicator size="small" />
+                            ) : (
+                              <ThemedText type="link" themeColor="textSecondary">
+                                {meta ? 'Looked up' : 'Look up'}
+                              </ThemedText>
+                            )}
+                          </Pressable>
+                        </View>
+
+                        {meta && (
+                          <View style={styles.metaBox}>
+                            {meta.title && <ThemedText type="smallBold">{meta.title}</ThemedText>}
+                            {(meta.authors?.length || meta.year) && (
+                              <ThemedText type="small" themeColor="textSecondary">
+                                {[meta.authors?.join(', '), meta.year].filter(Boolean).join(' · ')}
+                              </ThemedText>
+                            )}
+                            <ThemedText type="small" themeColor="textSecondary">{meta.source}</ThemedText>
+                          </View>
+                        )}
+
+                        <View style={styles.linkChips}>
+                          {linkEntries(links).map(([label, url]) => (
+                            <Pressable
+                              key={label}
+                              onPress={() => openLink(url)}
+                              style={({ pressed }) => [
+                                styles.linkChip,
+                                { backgroundColor: theme.backgroundSelected, opacity: pressed ? 0.7 : 1 },
+                              ]}>
+                              <ThemedText type="small">{label}</ThemedText>
+                            </Pressable>
+                          ))}
+                        </View>
+                      </View>
+                    );
+                  })
+                )}
+              </View>
+            </>
+          )}
+        </ThemedView>
+
         {segments.map((s, i) => (
-          <View key={i} style={[styles.segRow, i === activeIndex && { backgroundColor: theme.backgroundSelected }]}>
+          <View
+            key={i}
+            onLayout={(e) => {
+              segmentYs.current[i] = e.nativeEvent.layout.y;
+            }}
+            style={[styles.segRow, i === activeIndex && { backgroundColor: theme.backgroundSelected }]}>
             <Pressable onPress={() => seek(s.start)} hitSlop={6}>
               <ThemedText type="code" themeColor="textSecondary" style={styles.ts}>
                 {formatClock(s.start)}
@@ -148,7 +553,7 @@ export default function TranscriptScreen() {
             </Pressable>
             <TextInput
               value={s.text}
-              onChangeText={(t) => editSegment(i, t)}
+              onChangeText={(next) => editSegment(i, next)}
               multiline
               style={[styles.segText, { color: theme.text }]}
             />
@@ -172,10 +577,54 @@ export default function TranscriptScreen() {
   );
 }
 
+/** Index of the segment whose start is closest to `time` (fallback for jumps
+ *  that don't land inside any segment's [start, end) window). */
+function nearestIndex(segments: Segment[], time: number): number {
+  if (segments.length === 0) return -1;
+  let best = 0;
+  let bestDist = Infinity;
+  for (let i = 0; i < segments.length; i++) {
+    const seg = segments[i];
+    if (!seg) continue;
+    const dist = Math.abs(seg.start - time);
+    if (dist < bestDist) {
+      bestDist = dist;
+      best = i;
+    }
+  }
+  return best;
+}
+
 function Centered({ children }: { children: React.ReactNode }) {
   return <ThemedView style={[styles.fill, styles.centered]}>{children}</ThemedView>;
 }
 
+/** Localized date for a candidate event — date-only when it's an all-day item. */
+function formatEventDate(e: CandidateEvent): string {
+  const d = new Date(e.dateMs);
+  return e.allDay
+    ? d.toLocaleDateString()
+    : d.toLocaleString([], { dateStyle: 'medium', timeStyle: 'short' });
+}
+
+/** Ordered, labeled, defined link entries from a LookupLinks bag. */
+function linkEntries(links: LookupLinks): [string, string][] {
+  const order: [keyof LookupLinks, string][] = [
+    ['doi', 'DOI'],
+    ['arxiv', 'arXiv'],
+    ['openAlex', 'OpenAlex'],
+    ['openLibrary', 'Open Library'],
+    ['googleBooks', 'Google Books'],
+    ['wikipedia', 'Wikipedia'],
+    ['annasArchive', "Anna's Archive"],
+    ['libgen', 'LibGen'],
+  ];
+  return order
+    .map(([key, label]) => [label, links[key]] as const)
+    .filter((pair): pair is [string, string] => typeof pair[1] === 'string' && pair[1].length > 0)
+    .map(([label, url]) => [label, url]);
+}
+
 const styles = StyleSheet.create({
   fill: { flex: 1 },
   centered: { alignItems: 'center', justifyContent: 'center' },
@@ -183,6 +632,20 @@ const styles = StyleSheet.create({
   title: { fontSize: 22, fontWeight: '700', paddingVertical: Spacing.two },
   exportRow: { flexDirection: 'row', flexWrap: 'wrap', gap: Spacing.two, marginBottom: Spacing.two },
   chip: { paddingHorizontal: Spacing.three, paddingVertical: Spacing.one, borderRadius: 999 },
+  aidsCard: { padding: Spacing.three, borderRadius: Spacing.three, gap: Spacing.two, marginBottom: Spacing.two },
+  aidsRow: { flexDirection: 'row', flexWrap: 'wrap', gap: Spacing.two },
+  aidBtn: { paddingHorizontal: Spacing.three, paddingVertical: Spacing.two, borderRadius: Spacing.two, alignItems: 'center' },
+  aidBtnText: { color: '#fff', fontWeight: '700', fontSize: 14 },
+  aidBtnTextAlt: { fontWeight: '700', fontSize: 14 },
+  aidSection: { gap: Spacing.one, marginTop: Spacing.one },
+  aidLine: { flexDirection: 'row', gap: Spacing.two, alignItems: 'flex-start', paddingVertical: 2 },
+  flex: { flex: 1 },
+  refHeader: { flexDirection: 'row', alignItems: 'center', justifyContent: 'space-between', gap: Spacing.two },
+  refRow: { flexDirection: 'row', gap: Spacing.two, alignItems: 'flex-start', paddingVertical: 2 },
+  citCard: { gap: Spacing.one, paddingVertical: Spacing.one },
+  metaBox: { gap: 2, paddingLeft: Spacing.two },
+  linkChips: { flexDirection: 'row', flexWrap: 'wrap', gap: Spacing.one },
+  linkChip: { paddingHorizontal: Spacing.two, paddingVertical: 2, borderRadius: 999 },
   segRow: { flexDirection: 'row', gap: Spacing.two, paddingVertical: Spacing.one, paddingHorizontal: Spacing.two, borderRadius: Spacing.two, alignItems: 'flex-start' },
   ts: { paddingTop: 4, minWidth: 52 },
   segText: { flex: 1, fontSize: 16, lineHeight: 24, padding: 0 },

src/components/PreCaptureSheet.tsx

@@ -0,0 +1,171 @@
+// Modal shown before a transcription starts: capture title + course + lecture
+// date + language so the recording lands organized (ROADMAP Phase 0).
+import { useEffect, useState } from 'react';
+import { Modal, Pressable, ScrollView, StyleSheet, TextInput, View } from 'react-native';
+
+import { ThemedText } from '@/components/themed-text';
+import { ThemedView } from '@/components/themed-view';
+import { Spacing } from '@/constants/theme';
+import { useTheme } from '@/hooks/use-theme';
+import { useCourses } from '@/stores/coursesStore';
+
+export interface PreCaptureResult {
+  title: string;
+  courseId: string | null;
+  lectureDate?: number;
+  language?: string;
+}
+
+interface Props {
+  visible: boolean;
+  defaultTitle: string;
+  onStart: (r: PreCaptureResult) => void;
+  onCancel: () => void;
+}
+
+function todayYmd(): string {
+  const d = new Date();
+  const pad = (n: number) => String(n).padStart(2, '0');
+  return `${d.getFullYear()}-${pad(d.getMonth() + 1)}-${pad(d.getDate())}`;
+}
+
+function parseYmd(s: string): number | undefined {
+  const m = /^(\d{4})-(\d{2})-(\d{2})$/.exec(s.trim());
+  if (!m) return undefined;
+  const t = new Date(Number(m[1]), Number(m[2]) - 1, Number(m[3])).getTime();
+  return Number.isFinite(t) ? t : undefined;
+}
+
+export function PreCaptureSheet({ visible, defaultTitle, onStart, onCancel }: Props) {
+  const theme = useTheme();
+  const courses = useCourses((s) => s.items);
+  const refresh = useCourses((s) => s.refresh);
+  const createCourse = useCourses((s) => s.createCourse);
+
+  const [title, setTitle] = useState(defaultTitle);
+  const [courseId, setCourseId] = useState<string | null>(null);
+  const [dateStr, setDateStr] = useState(todayYmd());
+  const [language, setLanguage] = useState('');
+  const [newCourse, setNewCourse] = useState('');
+
+  useEffect(() => {
+    if (visible) {
+      setTitle(defaultTitle);
+      void refresh();
+    }
+  }, [visible, defaultTitle, refresh]);
+
+  const addCourse = async () => {
+    const name = newCourse.trim();
+    if (!name) return;
+    const c = await createCourse({ name });
+    setCourseId(c.id);
+    setNewCourse('');
+  };
+
+  const inputStyle = [styles.input, { color: theme.text, backgroundColor: theme.backgroundElement }];
+
+  return (
+    <Modal visible={visible} transparent animationType="slide" onRequestClose={onCancel}>
+      <View style={styles.backdrop}>
+        <ThemedView style={styles.sheet}>
+          <ScrollView contentContainerStyle={styles.content} keyboardShouldPersistTaps="handled">
+            <ThemedText type="subtitle">New transcription</ThemedText>
+
+            <ThemedText type="small" themeColor="textSecondary">Title</ThemedText>
+            <TextInput
+              value={title}
+              onChangeText={setTitle}
+              placeholder="Title"
+              placeholderTextColor={theme.textSecondary}
+              style={inputStyle}
+            />
+
+            <ThemedText type="small" themeColor="textSecondary">Course</ThemedText>
+            <View style={styles.chips}>
+              <Pressable
+                onPress={() => setCourseId(null)}
+                style={[styles.chip, { backgroundColor: courseId === null ? '#3c87f7' : theme.backgroundElement }]}>
+                <ThemedText type="small" style={courseId === null ? styles.chipActive : undefined}>Unsorted</ThemedText>
+              </Pressable>
+              {courses.map((c) => (
+                <Pressable
+                  key={c.id}
+                  onPress={() => setCourseId(c.id)}
+                  style={[styles.chip, { backgroundColor: courseId === c.id ? '#3c87f7' : theme.backgroundElement }]}>
+                  <ThemedText type="small" style={courseId === c.id ? styles.chipActive : undefined}>{c.name}</ThemedText>
+                </Pressable>
+              ))}
+            </View>
+            <View style={styles.row}>
+              <TextInput
+                value={newCourse}
+                onChangeText={setNewCourse}
+                placeholder="+ New course name"
+                placeholderTextColor={theme.textSecondary}
+                style={[inputStyle, styles.flex]}
+              />
+              <Pressable onPress={() => void addCourse()} style={[styles.addBtn, { opacity: newCourse.trim() ? 1 : 0.5 }]}>
+                <ThemedText type="small" style={styles.chipActive}>Add</ThemedText>
+              </Pressable>
+            </View>
+
+            <ThemedText type="small" themeColor="textSecondary">Lecture date</ThemedText>
+            <TextInput
+              value={dateStr}
+              onChangeText={setDateStr}
+              placeholder="YYYY-MM-DD"
+              placeholderTextColor={theme.textSecondary}
+              autoCapitalize="none"
+              style={inputStyle}
+            />
+
+            <ThemedText type="small" themeColor="textSecondary">Language (blank = auto-detect)</ThemedText>
+            <TextInput
+              value={language}
+              onChangeText={setLanguage}
+              placeholder="e.g. en, de"
+              placeholderTextColor={theme.textSecondary}
+              autoCapitalize="none"
+              style={inputStyle}
+            />
+
+            <View style={styles.actions}>
+              <Pressable onPress={onCancel} style={[styles.btn, { backgroundColor: theme.backgroundElement }]}>
+                <ThemedText>Cancel</ThemedText>
+              </Pressable>
+              <Pressable
+                onPress={() =>
+                  onStart({
+                    title: title.trim() || defaultTitle,
+                    courseId,
+                    lectureDate: parseYmd(dateStr),
+                    language: language.trim() || undefined,
+                  })
+                }
+                style={[styles.btn, { backgroundColor: '#3c87f7' }]}>
+                <ThemedText style={styles.startText}>Start</ThemedText>
+              </Pressable>
+            </View>
+          </ScrollView>
+        </ThemedView>
+      </View>
+    </Modal>
+  );
+}
+
+const styles = StyleSheet.create({
+  backdrop: { flex: 1, justifyContent: 'flex-end', backgroundColor: '#00000066' },
+  sheet: { maxHeight: '88%', borderTopLeftRadius: Spacing.four, borderTopRightRadius: Spacing.four },
+  content: { padding: Spacing.four, gap: Spacing.two, maxWidth: 800, width: '100%', alignSelf: 'center' },
+  input: { borderRadius: Spacing.two, paddingHorizontal: Spacing.three, paddingVertical: Spacing.two, fontSize: 15 },
+  chips: { flexDirection: 'row', flexWrap: 'wrap', gap: Spacing.two },
+  chip: { paddingHorizontal: Spacing.three, paddingVertical: Spacing.one, borderRadius: 999 },
+  chipActive: { color: '#fff', fontWeight: '700' },
+  row: { flexDirection: 'row', gap: Spacing.two, alignItems: 'center' },
+  flex: { flex: 1 },
+  addBtn: { backgroundColor: '#3c87f7', paddingHorizontal: Spacing.three, paddingVertical: Spacing.two, borderRadius: Spacing.two },
+  actions: { flexDirection: 'row', gap: Spacing.two, marginTop: Spacing.three },
+  btn: { flex: 1, paddingVertical: Spacing.three, borderRadius: Spacing.three, alignItems: 'center' },
+  startText: { color: '#fff', fontWeight: '700' },
+});

src/hooks/useRecorder.ts

@@ -0,0 +1,185 @@
+// Web microphone recorder (MediaRecorder) for in-app lecture capture (ROADMAP
+// Phase 2). Records to a Blob, exposes elapsed time + a live input level, holds
+// a screen Wake Lock while recording, and supports pause/resume. On stop it
+// returns the recording as an ArrayBuffer that flows straight into the existing
+// transcribe pipeline (decode -> chunk -> Whisper -> save).
+//
+// Web-only: getUserMedia/MediaRecorder/AudioContext are browser APIs. On native
+// `supported` is false and the controls no-op (native realtime capture via
+// whisper.rn is a follow-up).
+import { useCallback, useEffect, useRef, useState } from 'react';
+import { Platform } from 'react-native';
+
+export type RecorderStatus = 'idle' | 'recording' | 'paused' | 'stopped';
+
+export interface RecordingResult {
+  data: ArrayBuffer;
+  mime: string;
+  name: string;
+}
+
+export interface Recorder {
+  supported: boolean;
+  status: RecorderStatus;
+  elapsedMs: number;
+  /** Live input level in [0,1], for a meter. */
+  level: number;
+  error?: string;
+  start: () => Promise<void>;
+  pause: () => void;
+  resume: () => void;
+  /** Stop and return the recording (null if nothing captured / unsupported). */
+  stop: () => Promise<RecordingResult | null>;
+}
+
+function isSupported(): boolean {
+  return (
+    Platform.OS === 'web' &&
+    typeof navigator !== 'undefined' &&
+    !!navigator.mediaDevices?.getUserMedia &&
+    typeof MediaRecorder !== 'undefined'
+  );
+}
+
+export function useRecorder(): Recorder {
+  const supported = isSupported();
+  const [status, setStatus] = useState<RecorderStatus>('idle');
+  const [elapsedMs, setElapsedMs] = useState(0);
+  const [level, setLevel] = useState(0);
+  const [error, setError] = useState<string | undefined>(undefined);
+
+  const recorderRef = useRef<MediaRecorder | null>(null);
+  const chunksRef = useRef<Blob[]>([]);
+  const streamRef = useRef<MediaStream | null>(null);
+  const audioCtxRef = useRef<AudioContext | null>(null);
+  const analyserRef = useRef<AnalyserNode | null>(null);
+  const rafRef = useRef<number | null>(null);
+  const wakeRef = useRef<WakeLockSentinel | null>(null);
+  // Elapsed bookkeeping across pause/resume.
+  const segStartRef = useRef(0);
+  const accumRef = useRef(0);
+
+  const cleanup = useCallback(() => {
+    if (rafRef.current != null) cancelAnimationFrame(rafRef.current);
+    rafRef.current = null;
+    streamRef.current?.getTracks().forEach((t) => t.stop());
+    streamRef.current = null;
+    void audioCtxRef.current?.close().catch(() => {});
+    audioCtxRef.current = null;
+    analyserRef.current = null;
+    void wakeRef.current?.release().catch(() => {});
+    wakeRef.current = null;
+  }, []);
+
+  useEffect(() => cleanup, [cleanup]);
+
+  const tick = useCallback(() => {
+    // Elapsed: accumulated paused time + current running segment.
+    const running = status === 'recording' ? Date.now() - segStartRef.current : 0;
+    setElapsedMs(accumRef.current + running);
+
+    const an = analyserRef.current;
+    if (an) {
+      const buf = new Uint8Array(an.fftSize);
+      an.getByteTimeDomainData(buf);
+      let sum = 0;
+      for (let i = 0; i < buf.length; i++) {
+        const v = (buf[i]! - 128) / 128;
+        sum += v * v;
+      }
+      setLevel(Math.min(1, Math.sqrt(sum / buf.length) * 3));
+    }
+    rafRef.current = requestAnimationFrame(tick);
+  }, [status]);
+
+  // Keep the rAF loop bound to the latest `status` via tick's identity.
+  useEffect(() => {
+    if (status === 'recording' || status === 'paused') {
+      rafRef.current = requestAnimationFrame(tick);
+      return () => {
+        if (rafRef.current != null) cancelAnimationFrame(rafRef.current);
+      };
+    }
+    return undefined;
+  }, [status, tick]);
+
+  const start = useCallback(async () => {
+    if (!supported) {
+      setError('Recording is only available in the browser for now.');
+      return;
+    }
+    try {
+      setError(undefined);
+      const stream = await navigator.mediaDevices.getUserMedia({ audio: true });
+      streamRef.current = stream;
+
+      const mr = new MediaRecorder(stream);
+      chunksRef.current = [];
+      mr.ondataavailable = (e) => {
+        if (e.data && e.data.size > 0) chunksRef.current.push(e.data);
+      };
+      mr.start();
+      recorderRef.current = mr;
+
+      // Level metering off an AnalyserNode.
+      const ctx = new (window.AudioContext || (window as unknown as { webkitAudioContext: typeof AudioContext }).webkitAudioContext)();
+      const src = ctx.createMediaStreamSource(stream);
+      const an = ctx.createAnalyser();
+      an.fftSize = 512;
+      src.connect(an);
+      audioCtxRef.current = ctx;
+      analyserRef.current = an;
+
+      // Keep the screen awake while recording (best-effort).
+      try {
+        wakeRef.current = (await navigator.wakeLock?.request('screen')) ?? null;
+      } catch {
+        /* wake lock optional */
+      }
+
+      accumRef.current = 0;
+      segStartRef.current = Date.now();
+      setElapsedMs(0);
+      setStatus('recording');
+    } catch (e) {
+      setError(e instanceof Error ? e.message : 'Could not start recording (mic permission?).');
+      cleanup();
+      setStatus('idle');
+    }
+  }, [supported, cleanup]);
+
+  const pause = useCallback(() => {
+    const mr = recorderRef.current;
+    if (!mr || status !== 'recording') return;
+    mr.pause();
+    accumRef.current += Date.now() - segStartRef.current;
+    setStatus('paused');
+  }, [status]);
+
+  const resume = useCallback(() => {
+    const mr = recorderRef.current;
+    if (!mr || status !== 'paused') return;
+    mr.resume();
+    segStartRef.current = Date.now();
+    setStatus('recording');
+  }, [status]);
+
+  const stop = useCallback(async (): Promise<RecordingResult | null> => {
+    const mr = recorderRef.current;
+    if (!mr) return null;
+    const mime = mr.mimeType || 'audio/webm';
+    const blob: Blob = await new Promise((resolve) => {
+      mr.onstop = () => resolve(new Blob(chunksRef.current, { type: mime }));
+      mr.stop();
+    });
+    recorderRef.current = null;
+    cleanup();
+    setStatus('stopped');
+    if (blob.size === 0) return null;
+    const data = await blob.arrayBuffer();
+    const ext = mime.includes('ogg') ? 'ogg' : mime.includes('mp4') ? 'mp4' : 'webm';
+    return { data, mime, name: `Recording.${ext}` };
+  }, [cleanup]);
+
+  return { supported, status, elapsedMs, level, error, start, pause, resume, stop };
+}

src/lib/audio/decode.native.ts

@@ -1,27 +1,30 @@
 // Native (iOS/Android) audio decoder.
 //
-// Unlike the web, React Native has no built-in media decoder we can rely on.
+// WAV is decoded in pure JS (`decodeWav`) — fast and dependency-free. Every
-// `ffmpeg-kit-react-native` — the package everyone used to reach for — was
+// other container/codec (m4a/aac, mp3, ogg/opus, flac, …) is decoded by
-// RETIRED by its maintainer in early 2025, so we deliberately do NOT depend on
+// `react-native-audio-api`, which uses the platform codecs (Android MediaCodec /
-// it. For now we support only WAV, which we can decode in pure JS via our own
+// iOS AVFoundation, with a bundled FFmpeg fallback). This mirrors the web
-// `decodeWav`. Everything else throws a clear, actionable error.
+// decoder's `AudioContext.decodeAudioData` path. Decoding is fully on-device —
+// the audio bytes never leave the phone.
 //
-// TODO(audio/native): add a maintained native decoder for compressed formats
+// Why this matters on native: in-app recording is web-only, so importing a file
-//   (mp3/m4a/aac/ogg/flac). The current front-runner is
+// from the phone's own recorder is the ONLY way to get audio in — and those are
-//   `react-native-audio-api` (a Web-Audio-style API for RN). When added, route
+// almost always m4a/aac. (Previously native accepted WAV only.)
-//   non-WAV URIs through it and downmix/resample with `toMono16k`, mirroring the
+//
-//   web path.
+// We hand react-native-audio-api the raw file BYTES (ArrayBuffer), not a URI, so
+// we don't depend on how the OS file picker spells the path (file:// vs
+// content://) — `readFileBytes` already reads either via Expo's File API.
 //
 // File reading uses Expo SDK 56's object-oriented `File` API
 // (`new File(uri).bytes()`), which returns the raw bytes as a `Uint8Array`
-// directly — no base64 round-trip needed. We still keep a base64 decode helper
+// directly — no base64 round-trip needed. We keep a base64 fallback for
-// below as a documented fallback for environments/URIs where `.bytes()` isn't
+// environments/URIs where `.bytes()` isn't available.
-// available (it reads via `.base64()` and decodes to a Uint8Array by hand).
 
 import { File } from 'expo-file-system';
+import { decodeAudioData } from 'react-native-audio-api';
 
 import type { AudioDecoder, AudioFileInput } from './decode';
-import type { PcmAudio } from '../types';
+import { WHISPER_SAMPLE_RATE, type PcmAudio } from '../types';
 import { decodeWav } from './wav';
 import { toMono16k } from './resample';
 
@@ -90,6 +93,25 @@ async function readFileBytes(uri: string): Promise<Uint8Array> {
   }
 }
 
+/**
+ * True if these bytes are a RIFF/WAVE container. Sniffing the content (rather
+ * than trusting the file extension) lets us take the fast pure-JS WAV path even
+ * for mislabeled files, and avoids handing WAV to the heavier native decoder.
+ */
+function isWavBytes(b: Uint8Array): boolean {
+  return (
+    b.length >= 12 &&
+    b[0] === 0x52 && // 'R'
+    b[1] === 0x49 && // 'I'
+    b[2] === 0x46 && // 'F'
+    b[3] === 0x46 && // 'F'
+    b[8] === 0x57 && // 'W'
+    b[9] === 0x41 && // 'A'
+    b[10] === 0x56 && // 'V'
+    b[11] === 0x45 // 'E'
+  );
+}
+
 export const decoder: AudioDecoder = {
   async decode(input: AudioFileInput): Promise<PcmAudio> {
     const { uri } = input;
@@ -97,22 +119,42 @@ export const decoder: AudioDecoder = {
       throw new Error('Native decoder requires a file `uri`.');
     }
 
-    // Sniff the container from the URI extension (case-insensitive), ignoring
+    const bytes = await readFileBytes(uri);
-    // any query string / fragment that content URIs sometimes carry.
-    const path = uri.split(/[?#]/, 1)[0] ?? uri;
-    const isWav = path.toLowerCase().endsWith('.wav');
 
-    if (!isWav) {
+    // Fast path: WAV is decoded in pure JS (no native call).
+    if (isWavBytes(bytes)) {
+      const { sampleRate, channelData } = decodeWav(bytes);
+      return { sampleRate: 16000, samples: toMono16k(channelData, sampleRate) };
+    }
+
+    // Everything else (m4a/aac, mp3, ogg/opus, flac, …) -> platform codecs via
+    // react-native-audio-api. A fresh, contiguous ArrayBuffer of exactly the
+    // file bytes is what `decodeAudioData` expects.
+    const arrayBuffer = new Uint8Array(bytes).buffer as ArrayBuffer;
+    let audioBuffer;
+    try {
+      // Ask the native engine to resample to 16kHz DURING decode (platform DSP).
+      // This avoids materializing a full-rate (e.g. 44.1kHz stereo) Float32
+      // buffer in JS and skips the pure-JS resample loop — a big memory/CPU win
+      // and OOM guard on long lectures. toMono16k below still downmixes to mono
+      // and short-circuits its resampler since the rate already matches.
+      audioBuffer = await decodeAudioData(arrayBuffer, WHISPER_SAMPLE_RATE);
+    } catch (e) {
+      const detail = e instanceof Error ? e.message : String(e);
       throw new Error(
-        'Only WAV is supported on native for now — ffmpeg-kit-react-native ' +
+        `Couldn't decode this audio on device (${detail}). Supported formats: ` +
-          'was retired in 2025; a maintained native decoder (e.g. ' +
+          'WAV, m4a/aac, mp3, ogg, flac. If it keeps failing, convert it to ' +
-          'react-native-audio-api) is a follow-up.',
+          'WAV and import again.',
       );
     }
 
-    const bytes = await readFileBytes(uri);
+    const channelData: Float32Array[] = Array.from(
-    const { sampleRate, channelData } = decodeWav(bytes);
+      { length: audioBuffer.numberOfChannels },
-    const samples = toMono16k(channelData, sampleRate);
+      (_unused, c) => audioBuffer.getChannelData(c),
-    return { sampleRate: 16000, samples };
+    );
+    return {
+      sampleRate: 16000,
+      samples: toMono16k(channelData, audioBuffer.sampleRate),
+    };
   },
 };

src/lib/audio/wav.test.ts

@@ -197,6 +197,62 @@ describe('decodeWav stereo de-interleaving', () => {
     expect(decoded.channelData[0]![0]!).toBeCloseTo(12345 / 0x7fff, 4);
   });
 
+  it('decodes 24-bit PCM (format 1) data', () => {
+    // mono, 2 samples: +0x400000 (~+0.5) and -0x400000 (-0.5), 3 bytes LE each.
+    const data = new Uint8Array([0x00, 0x00, 0x40, 0x00, 0x00, 0xc0]);
+    const wav = buildWav({
+      audioFormat: 1,
+      numChannels: 1,
+      sampleRate: 16000,
+      bitsPerSample: 24,
+      data,
+    });
+    const decoded = decodeWav(wav);
+    expect(decoded.channelData.length).toBe(1);
+    expect(decoded.channelData[0]![0]!).toBeCloseTo(0x400000 / 0x7fffff, 3);
+    expect(decoded.channelData[0]![1]!).toBeCloseTo(-0.5, 3);
+  });
+
+  it('decodes WAVE_FORMAT_EXTENSIBLE (0xFFFE) wrapping 16-bit PCM', () => {
+    // 40-byte fmt: 16 standard + 24-byte extension whose SubFormat GUID begins
+    // with the real format tag (1 = PCM).
+    const fmt = new Uint8Array(40);
+    const fv = new DataView(fmt.buffer);
+    fv.setUint16(0, 0xfffe, true); // WAVE_FORMAT_EXTENSIBLE
+    fv.setUint16(2, 1, true); // channels
+    fv.setUint32(4, 16000, true); // sampleRate
+    fv.setUint32(8, 32000, true); // byteRate
+    fv.setUint16(12, 2, true); // blockAlign
+    fv.setUint16(14, 16, true); // bitsPerSample
+    fv.setUint16(16, 22, true); // cbSize
+    fv.setUint16(18, 16, true); // validBitsPerSample
+    fv.setUint32(20, 0, true); // channelMask
+    fv.setUint16(24, 1, true); // SubFormat GUID first 2 bytes -> PCM
+    const data = new Uint8Array(2);
+    new DataView(data.buffer).setInt16(0, 16383, true); // ~0.5
+
+    const total = 12 + (8 + 40) + (8 + 2);
+    const out = new Uint8Array(total);
+    const view = new DataView(out.buffer);
+    const tag = (off: number, s: string) => {
+      for (let i = 0; i < 4; i++) view.setUint8(off + i, s.charCodeAt(i));
+    };
+    tag(0, 'RIFF');
+    view.setUint32(4, total - 8, true);
+    tag(8, 'WAVE');
+    tag(12, 'fmt ');
+    view.setUint32(16, 40, true);
+    out.set(fmt, 20);
+    tag(60, 'data');
+    view.setUint32(64, 2, true);
+    out.set(data, 68);
+
+    const decoded = decodeWav(out);
+    expect(decoded.sampleRate).toBe(16000);
+    expect(decoded.channelData.length).toBe(1);
+    expect(decoded.channelData[0]![0]!).toBeCloseTo(16383 / 0x7fff, 4);
+  });
+
   it('decodes 32-bit IEEE float (format 3) data', () => {
     // mono, 2 samples: 0.25 and -0.75
     const data = new Uint8Array(8);

src/lib/audio/wav.ts

@@ -20,6 +20,10 @@ export interface DecodedWav {
 
 const WAVE_FORMAT_PCM = 1;
 const WAVE_FORMAT_IEEE_FLOAT = 3;
+// Many encoders (ffmpeg, Audacity, Windows recorders) wrap PCM/float in
+// WAVE_FORMAT_EXTENSIBLE; the real format tag is the first 2 bytes of the
+// extension's SubFormat GUID.
+const WAVE_FORMAT_EXTENSIBLE = 0xfffe;
 
 /** Read a 4-byte ASCII tag (e.g. "RIFF") at the given byte offset. */
 function readTag(view: DataView, offset: number): string {
@@ -81,6 +85,12 @@ export function decodeWav(bytes: Uint8Array): DecodedWav {
       sampleRate = view.getUint32(chunkBody + 4, true);
       // bytes 8..12: byteRate, bytes 12..14: blockAlign (derived; ignored)
       bitsPerSample = view.getUint16(chunkBody + 14, true);
+      // WAVE_FORMAT_EXTENSIBLE: the actual PCM/float tag lives in the extension's
+      // SubFormat GUID (first 2 bytes), at fmt-body + 24. Resolve it so common
+      // ffmpeg/Windows EXTENSIBLE WAVs decode instead of throwing.
+      if (audioFormat === WAVE_FORMAT_EXTENSIBLE && chunkBody + 26 <= bytes.byteLength) {
+        audioFormat = view.getUint16(chunkBody + 24, true);
+      }
       haveFmt = true;
     } else if (chunkId === 'data') {
       dataOffset = chunkBody;
@@ -124,6 +134,26 @@ export function decodeWav(bytes: Uint8Array): DecodedWav {
         channelData[c]![frame] = f;
       }
     }
+  } else if (audioFormat === WAVE_FORMAT_PCM && bitsPerSample === 24) {
+    // 24-bit signed little-endian PCM (common in EXTENSIBLE WAVs).
+    const bytesPerSample = 3;
+    const frameSize = bytesPerSample * numChannels;
+    const numFrames = Math.floor(dataLength / frameSize);
+    for (let c = 0; c < numChannels; c++) {
+      channelData.push(new Float32Array(numFrames));
+    }
+    for (let frame = 0; frame < numFrames; frame++) {
+      const base = dataOffset + frame * frameSize;
+      for (let c = 0; c < numChannels; c++) {
+        const o = base + c * bytesPerSample;
+        let v =
+          view.getUint8(o) |
+          (view.getUint8(o + 1) << 8) |
+          (view.getUint8(o + 2) << 16);
+        if (v & 0x800000) v -= 0x1000000; // sign-extend the 24th bit
+        channelData[c]![frame] = v < 0 ? v / 0x800000 : v / 0x7fffff;
+      }
+    }
   } else if (
     audioFormat === WAVE_FORMAT_IEEE_FLOAT &&
     bitsPerSample === 32
@@ -146,7 +176,8 @@ export function decodeWav(bytes: Uint8Array): DecodedWav {
   } else {
     throw new Error(
       `Unsupported WAV format: audioFormat=${audioFormat}, ` +
-        `bitsPerSample=${bitsPerSample} (supported: 16-bit PCM, 32-bit float)`,
+        `bitsPerSample=${bitsPerSample} ` +
+        `(supported: 16/24-bit PCM, 32-bit float, incl. WAVE_FORMAT_EXTENSIBLE)`,
     );
   }
 

src/lib/db/repo.test.ts

@@ -4,11 +4,59 @@
 // under vitest. We test the Dexie repo as the reference implementation of the
 // StorageRepo contract; the native expo-sqlite repo (native-only) is NOT
 // imported here because expo-sqlite cannot load outside a device/simulator.
+//
+// IMPORTANT: repo.web.ts is *not* statically imported. To exercise the v1->v2
+// migration we must first seed a v1-shaped DB with a raw Dexie connection, and
+// only THEN import repo.web (whose module-eval opens the DB at v2, triggering
+// the upgrade). The dynamically-imported repo is stashed in a module-level
+// binding so all subsequent tests reuse it.
 import 'fake-indexeddb/auto';
 
-import { describe, it, expect, beforeEach } from 'vitest';
+import Dexie from 'dexie';
-import { repo } from './repo.web';
+import { describe, it, expect, beforeAll, beforeEach } from 'vitest';
-import { parseDraft, type TranscriptDraft } from './schema';
+import {
+  parseDraft,
+  type TranscriptDraft,
+  type StorageRepo,
+  type SegmentVector,
+} from './index';
+
+// Populated by the migration beforeAll (which seeds v1 then imports repo.web).
+let repo: StorageRepo;
+
+// The v1 record we seed: NO segment ids, NO courseId, but WITH a searchText
+// column (exactly the legacy shape the v1 repo persisted).
+const V1_ID = 't_legacy';
+const v1Row = {
+  id: V1_ID,
+  title: 'Legacy Lecture',
+  createdAt: 1000,
+  updatedAt: 1000,
+  durationSec: 42.5,
+  modelId: 'tiny.en' as const,
+  segmentCount: 2,
+  segments: [
+    { start: 0, end: 2, text: 'legacy hello' },
+    { start: 2, end: 5, text: 'legacy world', confidence: 0.8 },
+  ],
+  searchText: 'legacy lecture\nlegacy hello\nlegacy world',
+};
+
+beforeAll(async () => {
+  // 1) Seed a v1 DB with a raw Dexie connection (only the v1 store/schema).
+  const v1 = new Dexie('wisp');
+  v1.version(1).stores({ transcripts: 'id, createdAt' });
+  await v1.open();
+  await v1.table('transcripts').put(v1Row);
+  v1.close();
+
+  // 2) Now import repo.web, which opens 'wisp' at v2 and runs the upgrade.
+  const mod = await import('./repo.web');
+  repo = mod.repo;
+
+  // Stub object-URL creation so getMediaUrl returns a string under Node.
+  globalThis.URL.createObjectURL = () => 'blob:mock';
+});
 
 // Helper: a minimal valid draft, overridable per test.
 function makeDraft(over: Partial<TranscriptDraft> = {}): TranscriptDraft {
@@ -24,31 +72,76 @@ function makeDraft(over: Partial<TranscriptDraft> = {}): TranscriptDraft {
   };
 }
 
-// Wipe storage between tests so ordering/search assertions are deterministic.
+describe('v1 -> v2 migration (no data loss)', () => {
-beforeEach(async () => {
+  it('keeps the legacy transcript, backfills segment ids, leaves it Unsorted', async () => {
-  const all = await repo.list();
+    const fetched = await repo.get(V1_ID);
-  await Promise.all(all.map((m) => repo.remove(m.id)));
+    expect(fetched).toBeDefined();
+
+    // Identity-preserving fields survive untouched.
+    expect(fetched!.title).toBe('Legacy Lecture');
+    expect(fetched!.durationSec).toBe(42.5);
+
+    // Segment text/timing is identical to the v1 body...
+    expect(fetched!.segments.map((s) => s.text)).toEqual([
+      'legacy hello',
+      'legacy world',
+    ]);
+    expect(fetched!.segments.map((s) => ({ start: s.start, end: s.end }))).toEqual([
+      { start: 0, end: 2 },
+      { start: 2, end: 5 },
+    ]);
+    expect(fetched!.segments[1]!.confidence).toBe(0.8);
+
+    // ...but every segment now carries a backfilled stable id.
+    for (const s of fetched!.segments) {
+      expect(s.id).toMatch(/^s_/);
+    }
+
+    // No courseId was set by v1 => the row is "Unsorted" (null/undefined).
+    expect(fetched!.courseId == null).toBe(true);
+
+    // And it shows up under the Unsorted bucket.
+    const unsorted = await repo.listByCourse(null);
+    expect(unsorted.map((m) => m.id)).toContain(V1_ID);
+  });
 });
 
 describe('StorageRepo (Dexie web impl)', () => {
-  it('create -> get round-trips segments and metadata', async () => {
+  // Wipe transcripts/courses/media between tests so ordering/search/course
+  // assertions are deterministic. (Runs after the migration beforeAll.)
+  beforeEach(async () => {
+    const all = await repo.list();
+    await Promise.all(all.map((m) => repo.remove(m.id)));
+    const courses = await repo.listCourses();
+    await Promise.all(courses.map((c) => repo.deleteCourse(c.id)));
+  });
+
+  it('create -> get round-trips segments and metadata, assigns segment ids', async () => {
     const created = await repo.create(makeDraft());
 
     expect(created.id).toMatch(/^t_/);
     expect(created.createdAt).toBeTypeOf('number');
     expect(created.updatedAt).toBe(created.createdAt);
     expect(created.segmentCount).toBe(2);
+    // courseId defaults to null ("Unsorted") when absent.
+    expect(created.courseId).toBeNull();
+    expect(created.hasMedia).toBe(false);
+    // Every segment got a stable id assigned on write.
+    for (const s of created.segments) {
+      expect(s.id).toMatch(/^s_/);
+    }
 
     const fetched = await repo.get(created.id);
     expect(fetched).toBeDefined();
     expect(fetched!.title).toBe('My Recording');
     expect(fetched!.durationSec).toBe(12.5);
     expect(fetched!.modelId).toBe('tiny.en');
-    // Segments survive the round-trip exactly.
+    // Text/timing survive the round-trip exactly.
-    expect(fetched!.segments).toEqual([
+    expect(fetched!.segments.map((s) => ({ start: s.start, end: s.end, text: s.text }))).toEqual([
       { start: 0, end: 2, text: 'hello world' },
-      { start: 2, end: 5, text: 'second segment', confidence: 0.9 },
+      { start: 2, end: 5, text: 'second segment' },
     ]);
+    expect(fetched!.segments[1]!.confidence).toBe(0.9);
   });
 
   it('get returns undefined for an unknown id', async () => {
@@ -57,7 +150,6 @@ describe('StorageRepo (Dexie web impl)', () => {
 
   it('list returns metadatas newest-first (without segments)', async () => {
     const a = await repo.create(makeDraft({ title: 'first' }));
-    // Force distinct, increasing createdAt timestamps.
     await new Promise((r) => setTimeout(r, 2));
     const b = await repo.create(makeDraft({ title: 'second' }));
     await new Promise((r) => setTimeout(r, 2));
@@ -65,12 +157,12 @@ describe('StorageRepo (Dexie web impl)', () => {
 
     const metas = await repo.list();
     expect(metas.map((m) => m.id)).toEqual([c.id, b.id, a.id]);
-    // Metas omit the segment body.
     expect((metas[0] as unknown as Record<string, unknown>).segments).toBeUndefined();
     expect(metas[0]!.segmentCount).toBe(2);
   });
 
-  it('update changes title + segments and re-derives searchText for search', async () => {
+  it('update merges title/segments/courseId and re-ids segments + re-derives search', async () => {
+    const course = await repo.createCourse({ name: 'Physics' });
     const created = await repo.create(
       makeDraft({ title: 'Old Title', segments: [{ start: 0, end: 1, text: 'apple' }] }),
     );
@@ -79,6 +171,7 @@ describe('StorageRepo (Dexie web impl)', () => {
 
     const updated = await repo.update(created.id, {
       title: 'Brand New Title',
+      courseId: course.id,
       segments: [
         { start: 0, end: 1, text: 'banana' },
         { start: 1, end: 2, text: 'cherry' },
@@ -86,16 +179,19 @@ describe('StorageRepo (Dexie web impl)', () => {
     });
 
     expect(updated.title).toBe('Brand New Title');
+    expect(updated.courseId).toBe(course.id);
     expect(updated.segmentCount).toBe(2);
     expect(updated.updatedAt).toBeGreaterThan(originalUpdatedAt);
+    // New segments got ids.
+    for (const s of updated.segments) {
+      expect(s.id).toMatch(/^s_/);
+    }
 
     // Old text is no longer findable...
     expect(await repo.search('apple')).toHaveLength(0);
     // ...and the new text + new title are.
-    const byText = await repo.search('cherry');
+    expect((await repo.search('cherry')).map((m) => m.id)).toEqual([created.id]);
-    expect(byText.map((m) => m.id)).toEqual([created.id]);
+    expect((await repo.search('brand new')).map((m) => m.id)).toEqual([created.id]);
-    const byTitle = await repo.search('brand new');
-    expect(byTitle.map((m) => m.id)).toEqual([created.id]);
   });
 
   it('update throws for an unknown id', async () => {
@@ -110,24 +206,106 @@ describe('StorageRepo (Dexie web impl)', () => {
       }),
     );
 
-    // Title match, different case.
     expect((await repo.search('meeting')).map((m) => m.id)).toEqual([t.id]);
-    // Segment-text match, different case.
     expect((await repo.search('budget')).map((m) => m.id)).toEqual([t.id]);
-    // Non-match.
     expect(await repo.search('zzz-not-present')).toHaveLength(0);
   });
 
-  it('remove deletes a transcript', async () => {
+  it('remove deletes a transcript (and its media)', async () => {
     const created = await repo.create(makeDraft());
+    await repo.putMedia(created.id, { data: new ArrayBuffer(4), mime: 'audio/mpeg' });
     expect(await repo.get(created.id)).toBeDefined();
 
     await repo.remove(created.id);
     expect(await repo.get(created.id)).toBeUndefined();
     expect(await repo.list()).toHaveLength(0);
+    // Media is gone too: no playable URL remains.
+    expect(await repo.getMediaUrl(created.id)).toBeUndefined();
   });
 
-  it('parseDraft rejects a NaN duration', () => {
+  // --- courses -------------------------------------------------------------
+
+  it('listByCourse splits Unsorted vs a course', async () => {
+    const course = await repo.createCourse({ name: 'Bio 101' });
+    const inCourse = await repo.create(makeDraft({ title: 'lecture', courseId: course.id }));
+    const unsorted = await repo.create(makeDraft({ title: 'loose note' }));
+
+    const courseRows = await repo.listByCourse(course.id);
+    expect(courseRows.map((m) => m.id)).toEqual([inCourse.id]);
+
+    const unsortedRows = await repo.listByCourse(null);
+    expect(unsortedRows.map((m) => m.id)).toEqual([unsorted.id]);
+  });
+
+  it('createCourse / listCourses (alphabetical) / updateCourse', async () => {
+    const zebra = await repo.createCourse({ name: 'Zebra Studies', code: 'ZOO-1' });
+    const apple = await repo.createCourse({ name: 'Apple Theory' });
+
+    expect(apple.id).toMatch(/^c_/);
+    expect(apple.createdAt).toBeTypeOf('number');
+
+    // Alphabetical by name.
+    const listed = await repo.listCourses();
+    expect(listed.map((c) => c.name)).toEqual(['Apple Theory', 'Zebra Studies']);
+
+    const renamed = await repo.updateCourse(zebra.id, { name: 'Aardvark Studies' });
+    expect(renamed.name).toBe('Aardvark Studies');
+    expect(renamed.code).toBe('ZOO-1'); // unspecified fields preserved
+    expect(renamed.updatedAt).toBeGreaterThanOrEqual(zebra.updatedAt);
+
+    const relisted = await repo.listCourses();
+    expect(relisted.map((c) => c.name)).toEqual(['Aardvark Studies', 'Apple Theory']);
+  });
+
+  it('deleteCourse reassigns its transcripts to Unsorted, then removes the course', async () => {
+    const course = await repo.createCourse({ name: 'Doomed' });
+    const t = await repo.create(makeDraft({ title: 'orphan-to-be', courseId: course.id }));
+
+    await repo.deleteCourse(course.id);
+
+    // The course is gone.
+    expect((await repo.listCourses()).map((c) => c.id)).not.toContain(course.id);
+    // The transcript survives, now Unsorted.
+    const fetched = await repo.get(t.id);
+    expect(fetched).toBeDefined();
+    expect(fetched!.courseId).toBeNull();
+    expect((await repo.listByCourse(null)).map((m) => m.id)).toContain(t.id);
+  });
+
+  it('reassign moves a transcript between courses and to Unsorted', async () => {
+    const course = await repo.createCourse({ name: 'Chemistry' });
+    const t = await repo.create(makeDraft());
+    expect(t.courseId).toBeNull();
+
+    await repo.reassign(t.id, course.id);
+    expect((await repo.listByCourse(course.id)).map((m) => m.id)).toEqual([t.id]);
+
+    await repo.reassign(t.id, null);
+    expect((await repo.listByCourse(course.id))).toHaveLength(0);
+    expect((await repo.listByCourse(null)).map((m) => m.id)).toContain(t.id);
+  });
+
+  // --- media ---------------------------------------------------------------
+
+  it('putMedia flips hasMedia true and getMediaUrl returns a url; removeMedia flips it false', async () => {
+    const t = await repo.create(makeDraft());
+    expect(t.hasMedia).toBe(false);
+
+    await repo.putMedia(t.id, { data: new ArrayBuffer(8), mime: 'audio/mpeg' });
+
+    const afterPut = await repo.get(t.id);
+    expect(afterPut!.hasMedia).toBe(true);
+    expect(await repo.getMediaUrl(t.id)).toBe('blob:mock');
+
+    await repo.removeMedia(t.id);
+    const afterRemove = await repo.get(t.id);
+    expect(afterRemove!.hasMedia).toBe(false);
+    expect(await repo.getMediaUrl(t.id)).toBeUndefined();
+  });
+
+  // --- validation gatekeeping ---------------------------------------------
+
+  it('parseDraft rejects a NaN/Infinity duration', () => {
     expect(() => parseDraft(makeDraft({ durationSec: NaN }))).toThrow();
     expect(() => parseDraft(makeDraft({ durationSec: Infinity }))).toThrow();
   });
@@ -142,9 +320,394 @@ describe('StorageRepo (Dexie web impl)', () => {
   });
 
   it('create rejects an invalid draft (defense in depth)', async () => {
-    // create() calls parseDraft internally, so bad input never persists.
+    await expect(repo.create(makeDraft({ durationSec: -1 }))).rejects.toThrow();
-    await expect(
+  });
-      repo.create(makeDraft({ durationSec: -1 })),
+});
-    ).rejects.toThrow();
+
+// ---------------------------------------------------------------------------
+// Phase 1: semantic-search vector store
+// ---------------------------------------------------------------------------
+//
+// We exercise the brute-force cosine store with tiny hand-made UNIT vectors so
+// the math is checkable by eye. The repo must not hard-code 384 dims — these
+// tests use 3-d vectors throughout. Cosine of two unit vectors is their dot
+// product, so [1,0,0]·[1,0,0] = 1 and [1,0,0]·[0,1,0] = 0.
+
+const MODEL = 'Xenova/multilingual-e5-small';
+
+// Three orthonormal basis vectors, used as stand-in segment embeddings.
+const E0 = () => new Float32Array([1, 0, 0]);
+const E1 = () => new Float32Array([0, 1, 0]);
+const E2 = () => new Float32Array([0, 0, 1]);
+
+describe('vectors (semantic search store)', () => {
+  beforeEach(async () => {
+    // Wipe transcripts (cascades drop their vectors) + courses between tests.
+    const all = await repo.list();
+    await Promise.all(all.map((m) => repo.remove(m.id)));
+    const courses = await repo.listCourses();
+    await Promise.all(courses.map((c) => repo.deleteCourse(c.id)));
+  });
+
+  // Create a transcript whose segments carry stable ids, then return it.
+  async function makeTranscript(over: Partial<TranscriptDraft> = {}) {
+    return repo.create(
+      makeDraft({
+        segments: [
+          { start: 0, end: 1, text: 'alpha' },
+          { start: 1, end: 2, text: 'beta' },
+          { start: 2, end: 3, text: 'gamma' },
+        ],
+        ...over,
+      }),
+    );
+  }
+
+  // Build SegmentVector rows pairing each segment id with one of the vectors.
+  function vecsFor(
+    t: { segments: { id?: string; start: number; end: number; text: string }[] },
+    vectors: Float32Array[],
+  ): SegmentVector[] {
+    return t.segments.map((s, i) => ({
+      segmentId: s.id!,
+      start: s.start,
+      end: s.end,
+      text: s.text,
+      vector: vectors[i]!,
+    }));
+  }
+
+  it('upsertVectors + searchVectors ranks the matching segment first with score ~1', async () => {
+    const t = await makeTranscript();
+    await repo.upsertVectors(t.id, MODEL, vecsFor(t, [E0(), E1(), E2()]));
+
+    const hits = await repo.searchVectors(E0());
+    expect(hits).toHaveLength(3);
+    // The [1,0,0] segment ("alpha") ranks first with cosine ~1.
+    expect(hits[0]!.segmentId).toBe(t.segments[0]!.id);
+    expect(hits[0]!.text).toBe('alpha');
+    expect(hits[0]!.score).toBeCloseTo(1, 6);
+    // The orthogonal segments score ~0.
+    expect(hits[1]!.score).toBeCloseTo(0, 6);
+    expect(hits[2]!.score).toBeCloseTo(0, 6);
+    // Hits carry the segment-jump anchors + denormalized courseId.
+    expect(hits[0]!.transcriptId).toBe(t.id);
+    expect(hits[0]!.start).toBe(0);
+    expect(hits[0]!.end).toBe(1);
+    expect(hits[0]!.courseId).toBeNull();
+  });
+
+  it('searchVectors honours the limit option', async () => {
+    const t = await makeTranscript();
+    await repo.upsertVectors(t.id, MODEL, vecsFor(t, [E0(), E1(), E2()]));
+
+    const hits = await repo.searchVectors(E0(), { limit: 1 });
+    expect(hits).toHaveLength(1);
+    expect(hits[0]!.segmentId).toBe(t.segments[0]!.id);
+  });
+
+  it('upsertVectors replaces prior vectors for a transcript (no stale rows)', async () => {
+    const t = await makeTranscript();
+    await repo.upsertVectors(t.id, MODEL, vecsFor(t, [E0(), E1(), E2()]));
+    // Re-embed with only the first two segments.
+    await repo.upsertVectors(t.id, MODEL, [
+      {
+        segmentId: t.segments[0]!.id!,
+        start: 0,
+        end: 1,
+        text: 'alpha',
+        vector: E0(),
+      },
+      {
+        segmentId: t.segments[1]!.id!,
+        start: 1,
+        end: 2,
+        text: 'beta',
+        vector: E1(),
+      },
+    ]);
+    const hits = await repo.searchVectors(E0());
+    expect(hits).toHaveLength(2);
+  });
+
+  it('courseId filter scopes search (course vs Unsorted vs all)', async () => {
+    const course = await repo.createCourse({ name: 'Physics' });
+    const inCourse = await makeTranscript({ courseId: course.id });
+    const unsorted = await makeTranscript();
+
+    await repo.upsertVectors(inCourse.id, MODEL, vecsFor(inCourse, [E0(), E1(), E2()]));
+    await repo.upsertVectors(unsorted.id, MODEL, vecsFor(unsorted, [E0(), E1(), E2()]));
+
+    // Scoped to the course: only that transcript's rows are searched.
+    const courseHits = await repo.searchVectors(E0(), { courseId: course.id });
+    expect(courseHits.every((h) => h.transcriptId === inCourse.id)).toBe(true);
+    expect(courseHits.every((h) => h.courseId === course.id)).toBe(true);
+    expect(courseHits).toHaveLength(3);
+
+    // Scoped to Unsorted (null): only the unsorted transcript's rows.
+    const unsortedHits = await repo.searchVectors(E0(), { courseId: null });
+    expect(unsortedHits.every((h) => h.transcriptId === unsorted.id)).toBe(true);
+    expect(unsortedHits.every((h) => h.courseId === null)).toBe(true);
+    expect(unsortedHits).toHaveLength(3);
+
+    // No scope => everything (both transcripts).
+    const allHits = await repo.searchVectors(E0());
+    expect(allHits).toHaveLength(6);
+  });
+
+  it('unembeddedIds excludes embedded transcripts and includes fresh ones', async () => {
+    const embedded = await makeTranscript();
+    const fresh = await makeTranscript();
+    await repo.upsertVectors(embedded.id, MODEL, vecsFor(embedded, [E0(), E1(), E2()]));
+
+    const pending = await repo.unembeddedIds(MODEL);
+    expect(pending).not.toContain(embedded.id);
+    expect(pending).toContain(fresh.id);
+
+    // A different model id has no vectors at all => both are unembedded.
+    const otherModel = await repo.unembeddedIds('some/other-model');
+    expect(otherModel).toContain(embedded.id);
+    expect(otherModel).toContain(fresh.id);
+  });
+
+  it('clearVectors empties a transcript and re-adds it to unembeddedIds', async () => {
+    const t = await makeTranscript();
+    await repo.upsertVectors(t.id, MODEL, vecsFor(t, [E0(), E1(), E2()]));
+    expect(await repo.searchVectors(E0())).toHaveLength(3);
+
+    await repo.clearVectors(t.id);
+    expect(await repo.searchVectors(E0())).toHaveLength(0);
+    expect(await repo.unembeddedIds(MODEL)).toContain(t.id);
+  });
+
+  it('reassign updates the denormalized courseId on existing hits', async () => {
+    const course = await repo.createCourse({ name: 'Chemistry' });
+    const t = await makeTranscript();
+    await repo.upsertVectors(t.id, MODEL, vecsFor(t, [E0(), E1(), E2()]));
+
+    // Initially Unsorted: hits carry courseId null and match the null scope.
+    expect((await repo.searchVectors(E0(), { courseId: null }))).toHaveLength(3);
+
+    await repo.reassign(t.id, course.id);
+
+    // Now the hits carry the new courseId and match the course scope...
+    const courseHits = await repo.searchVectors(E0(), { courseId: course.id });
+    expect(courseHits).toHaveLength(3);
+    expect(courseHits.every((h) => h.courseId === course.id)).toBe(true);
+    // ...and no longer appear under Unsorted.
+    expect(await repo.searchVectors(E0(), { courseId: null })).toHaveLength(0);
+  });
+
+  it('updating segments clears stale vectors (transcript becomes unembedded again)', async () => {
+    const t = await makeTranscript();
+    await repo.upsertVectors(t.id, MODEL, vecsFor(t, [E0(), E1(), E2()]));
+    expect(await repo.unembeddedIds(MODEL)).not.toContain(t.id);
+
+    // Patching segments invalidates the embeddings => repo drops them.
+    await repo.update(t.id, {
+      segments: [{ start: 0, end: 1, text: 'rewritten' }],
+    });
+
+    expect(await repo.searchVectors(E0())).toHaveLength(0);
+    expect(await repo.unembeddedIds(MODEL)).toContain(t.id);
+  });
+
+  it('remove cascades to a transcript vectors', async () => {
+    const t = await makeTranscript();
+    await repo.upsertVectors(t.id, MODEL, vecsFor(t, [E0(), E1(), E2()]));
+    expect(await repo.searchVectors(E0())).toHaveLength(3);
+
+    await repo.remove(t.id);
+    expect(await repo.searchVectors(E0())).toHaveLength(0);
+  });
+});
+
+// ---------------------------------------------------------------------------
+// Phase 3: flashcards + SM-2 spaced repetition store
+// ---------------------------------------------------------------------------
+//
+// These exercise the storage-only flashcard methods: a brand-new card is due
+// immediately (initial srs.due ~ now); listDueFlashcards honours the due cutoff
+// and course scope; updateFlashcardSrs persists a recomputed schedule; counts,
+// deletes, and the remove(transcript) cascade all behave.
+
+describe('flashcards', () => {
+  beforeEach(async () => {
+    // Wipe transcripts (cascades drop their flashcards) + courses between tests.
+    const all = await repo.list();
+    await Promise.all(all.map((m) => repo.remove(m.id)));
+    const courses = await repo.listCourses();
+    await Promise.all(courses.map((c) => repo.deleteCourse(c.id)));
+  });
+
+  it('createFlashcards returns ids + an initial srs due ~ now', async () => {
+    const t = await repo.create(makeDraft());
+    const before = Date.now();
+    const [card] = await repo.createFlashcards([
+      {
+        transcriptId: t.id,
+        front: 'What is Q?',
+        back: 'The answer.',
+        segmentId: t.segments[0]!.id,
+        start: 0,
+      },
+    ]);
+    const after = Date.now();
+
+    expect(card!.id).toMatch(/^f_/);
+    expect(card!.transcriptId).toBe(t.id);
+    expect(card!.front).toBe('What is Q?');
+    expect(card!.back).toBe('The answer.');
+    expect(card!.segmentId).toBe(t.segments[0]!.id);
+    expect(card!.start).toBe(0);
+    expect(card!.createdAt).toBeTypeOf('number');
+    // Initial SM-2 state.
+    expect(card!.srs.ease).toBe(2.5);
+    expect(card!.srs.intervalDays).toBe(0);
+    expect(card!.srs.reps).toBe(0);
+    expect(card!.srs.lapses).toBe(0);
+    // Due immediately: srs.due falls within the create() window.
+    expect(card!.srs.due).toBeGreaterThanOrEqual(before);
+    expect(card!.srs.due).toBeLessThanOrEqual(after);
+
+    // It round-trips through listFlashcards.
+    const listed = await repo.listFlashcards({ transcriptId: t.id });
+    expect(listed.map((c) => c.id)).toEqual([card!.id]);
+  });
+
+  it('listDueFlashcards includes a due card and excludes a future-due one', async () => {
+    const t = await repo.create(makeDraft());
+    const [dueCard, futureCard] = await repo.createFlashcards([
+      { transcriptId: t.id, front: 'due', back: 'now' },
+      { transcriptId: t.id, front: 'later', back: 'future' },
+    ]);
+
+    // Push the second card's due far into the future via a recomputed schedule.
+    const future = Date.now() + 7 * 24 * 60 * 60 * 1000;
+    await repo.updateFlashcardSrs(futureCard!.id, {
+      ease: 2.5,
+      intervalDays: 7,
+      reps: 1,
+      lapses: 0,
+      due: future,
+    });
+
+    const due = await repo.listDueFlashcards();
+    expect(due.map((c) => c.id)).toContain(dueCard!.id);
+    expect(due.map((c) => c.id)).not.toContain(futureCard!.id);
+  });
+
+  it('updateFlashcardSrs persists the recomputed schedule', async () => {
+    const t = await repo.create(makeDraft());
+    const [card] = await repo.createFlashcards([
+      { transcriptId: t.id, front: 'q', back: 'a' },
+    ]);
+
+    const reviewed = Date.now();
+    const next = {
+      ease: 2.6,
+      intervalDays: 1,
+      reps: 1,
+      lapses: 0,
+      due: reviewed + 24 * 60 * 60 * 1000,
+      lastReviewed: reviewed,
+    };
+    const updated = await repo.updateFlashcardSrs(card!.id, next);
+    expect(updated.srs).toEqual(next);
+
+    // The change is durable.
+    const [reloaded] = await repo.listFlashcards({ transcriptId: t.id });
+    expect(reloaded!.srs).toEqual(next);
+  });
+
+  it('updateFlashcardSrs throws for an unknown id', async () => {
+    await expect(
+      repo.updateFlashcardSrs('f_nope', {
+        ease: 2.5,
+        intervalDays: 0,
+        reps: 0,
+        lapses: 0,
+        due: Date.now(),
+      }),
+    ).rejects.toThrow();
+  });
+
+  it('flashcardCounts reports total and due', async () => {
+    const t = await repo.create(makeDraft());
+    const [, b, c] = await repo.createFlashcards([
+      { transcriptId: t.id, front: '1', back: 'a' },
+      { transcriptId: t.id, front: '2', back: 'b' },
+      { transcriptId: t.id, front: '3', back: 'c' },
+    ]);
+
+    // All three start due.
+    expect(await repo.flashcardCounts()).toEqual({ total: 3, due: 3 });
+
+    // Push two out into the future => only one remains due.
+    const future = Date.now() + 60 * 60 * 1000;
+    const futureSrs = {
+      ease: 2.5,
+      intervalDays: 1,
+      reps: 1,
+      lapses: 0,
+      due: future,
+    };
+    await repo.updateFlashcardSrs(b!.id, futureSrs);
+    await repo.updateFlashcardSrs(c!.id, futureSrs);
+
+    expect(await repo.flashcardCounts()).toEqual({ total: 3, due: 1 });
+  });
+
+  it('deleteFlashcard removes a single card', async () => {
+    const t = await repo.create(makeDraft());
+    const [a, b] = await repo.createFlashcards([
+      { transcriptId: t.id, front: '1', back: 'a' },
+      { transcriptId: t.id, front: '2', back: 'b' },
+    ]);
+
+    await repo.deleteFlashcard(a!.id);
+    const remaining = await repo.listFlashcards({ transcriptId: t.id });
+    expect(remaining.map((c) => c.id)).toEqual([b!.id]);
+  });
+
+  it('remove(transcript) cascades to its flashcards', async () => {
+    const t = await repo.create(makeDraft());
+    await repo.createFlashcards([
+      { transcriptId: t.id, front: '1', back: 'a' },
+      { transcriptId: t.id, front: '2', back: 'b' },
+    ]);
+    expect(await repo.listFlashcards({ transcriptId: t.id })).toHaveLength(2);
+
+    await repo.remove(t.id);
+    expect(await repo.listFlashcards({ transcriptId: t.id })).toHaveLength(0);
+    expect(await repo.flashcardCounts()).toEqual({ total: 0, due: 0 });
+  });
+
+  it('courseId scopes listing, due, and counts; deleteCourse moves cards to Unsorted', async () => {
+    const course = await repo.createCourse({ name: 'Physics' });
+    const inCourse = await repo.create(makeDraft({ courseId: course.id }));
+    const unsorted = await repo.create(makeDraft());
+
+    await repo.createFlashcards([
+      { transcriptId: inCourse.id, courseId: course.id, front: 'c1', back: 'a' },
+    ]);
+    await repo.createFlashcards([
+      { transcriptId: unsorted.id, front: 'u1', back: 'b' },
+    ]);
+
+    // Listing is scoped.
+    expect((await repo.listFlashcards({ courseId: course.id })).map((c) => c.front)).toEqual(['c1']);
+    expect((await repo.listFlashcards({ courseId: null })).map((c) => c.front)).toEqual(['u1']);
+
+    // Counts + due are scoped.
+    expect(await repo.flashcardCounts(course.id)).toEqual({ total: 1, due: 1 });
+    expect(await repo.flashcardCounts(null)).toEqual({ total: 1, due: 1 });
+    expect((await repo.listDueFlashcards({ courseId: course.id })).map((c) => c.front)).toEqual(['c1']);
+
+    // Deleting the course moves its cards to Unsorted (courseId=null).
+    await repo.deleteCourse(course.id);
+    const movedToUnsorted = await repo.listFlashcards({ courseId: null });
+    expect(movedToUnsorted.map((c) => c.front).sort()).toEqual(['c1', 'u1']);
+    expect(await repo.listFlashcards({ courseId: course.id })).toHaveLength(0);
   });
 });

src/lib/db/repo.ts

@@ -5,9 +5,51 @@
 // and never on a specific backend. The correct implementation is selected at
 // build time via platform file extensions (see repoImpl.web.ts / repoImpl.native.ts).
 
-import type { TranscriptDraft, TranscriptMeta, Transcript } from './schema';
+import type {
+  TranscriptDraft,
+  TranscriptMeta,
+  Transcript,
+  TranscriptPatch,
+  Course,
+  CourseDraft,
+  Flashcard,
+  FlashcardDraft,
+  SrsState,
+} from './schema';
+
+/** Source of audio to persist for a transcript (web passes data, native a uri). */
+export interface MediaInput {
+  /** Raw bytes (web import path). */
+  data?: ArrayBuffer;
+  /** A file uri to copy into app storage (native import path). */
+  uri?: string;
+  /** MIME type, e.g. 'audio/mpeg'. */
+  mime: string;
+}
+
+/** One stored embedding for a segment (vector is unit-normalized). */
+export interface SegmentVector {
+  segmentId: string;
+  start: number;
+  end: number;
+  text: string;
+  vector: Float32Array;
+}
+
+/** A semantic-search hit at segment granularity. */
+export interface VectorHit {
+  transcriptId: string;
+  segmentId: string;
+  start: number;
+  end: number;
+  courseId: string | null;
+  text: string;
+  /** Cosine similarity (dot product of unit vectors). */
+  score: number;
+}
 
 export interface StorageRepo {
+  // --- transcripts ---------------------------------------------------------
   /** All transcript metadatas, newest first (by createdAt desc). */
   list(): Promise<TranscriptMeta[]>;
 
@@ -18,16 +60,14 @@ export interface StorageRepo {
   create(draft: TranscriptDraft): Promise<Transcript>;
 
   /**
-   * Patch a transcript's title and/or segments. Re-derives any denormalized
+   * Patch a transcript's mutable fields (title/segments/course+lecture meta).
-   * fields (searchText, segmentCount, updatedAt) and re-validates segments.
+   * Re-derives denormalized fields (searchText, segmentCount, updatedAt),
+   * re-validates segments, and assigns ids to any segments missing one.
    * Rejects if the id does not exist.
    */
-  update(
+  update(id: string, patch: TranscriptPatch): Promise<Transcript>;
-    id: string,
-    patch: Partial<Pick<Transcript, 'title' | 'segments'>>,
-  ): Promise<Transcript>;
 
-  /** Delete by id. No-op if absent. */
+  /** Delete by id (also removes its persisted media). No-op if absent. */
   remove(id: string): Promise<void>;
 
   /**
@@ -35,4 +75,78 @@ export interface StorageRepo {
    * Returns metadatas, newest first.
    */
   search(query: string): Promise<TranscriptMeta[]>;
+
+  /** Transcripts in a course (pass null for "Unsorted"), newest first. */
+  listByCourse(courseId: string | null): Promise<TranscriptMeta[]>;
+
+  /** Move a transcript to a course (or null for Unsorted). */
+  reassign(transcriptId: string, courseId: string | null): Promise<void>;
+
+  // --- courses -------------------------------------------------------------
+  /** All courses, alphabetical by name. */
+  listCourses(): Promise<Course[]>;
+
+  /** Validate + create a course. */
+  createCourse(draft: CourseDraft): Promise<Course>;
+
+  /** Patch a course's fields. Rejects if absent. */
+  updateCourse(id: string, patch: Partial<CourseDraft>): Promise<Course>;
+
+  /** Delete a course; its transcripts are reassigned to Unsorted (courseId=null). */
+  deleteCourse(id: string): Promise<void>;
+
+  // --- media (persisted source audio) -------------------------------------
+  /** Store (or replace) the source audio for a transcript. */
+  putMedia(transcriptId: string, input: MediaInput): Promise<void>;
+
+  /**
+   * A playable URL for the persisted audio (object URL on web, file uri on
+   * native), or undefined if no media is stored. Callers should revoke web
+   * object URLs when done.
+   */
+  getMediaUrl(transcriptId: string): Promise<string | undefined>;
+
+  /** Delete the persisted audio for a transcript. No-op if absent. */
+  removeMedia(transcriptId: string): Promise<void>;
+
+  // --- semantic search vectors (Phase 1) ----------------------------------
+  /**
+   * Replace all stored vectors for a transcript with `vectors`, tagged with the
+   * embedding `model` id (so a model change can be detected and re-embedded).
+   */
+  upsertVectors(transcriptId: string, model: string, vectors: SegmentVector[]): Promise<void>;
+
+  /**
+   * Brute-force cosine search over stored vectors (optionally scoped to a
+   * course; null = Unsorted), returning the top `limit` segment hits.
+   */
+  searchVectors(
+    query: Float32Array,
+    opts?: { courseId?: string | null; limit?: number },
+  ): Promise<VectorHit[]>;
+
+  /** Drop all vectors for a transcript. */
+  clearVectors(transcriptId: string): Promise<void>;
+
+  /** Transcript ids that have NO vectors for `model` (need embedding/backfill). */
+  unembeddedIds(model: string): Promise<string[]>;
+
+  // --- flashcards + spaced repetition (Phase 3) ---------------------------
+  /** Validate + persist new flashcards (each gets an id, createdAt, initial SRS). */
+  createFlashcards(drafts: FlashcardDraft[]): Promise<Flashcard[]>;
+
+  /** All flashcards, optionally scoped to a transcript or course. */
+  listFlashcards(opts?: { transcriptId?: string; courseId?: string | null }): Promise<Flashcard[]>;
+
+  /** Cards due for review (srs.due <= now), optionally course-scoped. */
+  listDueFlashcards(opts?: { courseId?: string | null; now?: number; limit?: number }): Promise<Flashcard[]>;
+
+  /** Persist a recomputed SRS schedule after a review. */
+  updateFlashcardSrs(id: string, srs: SrsState): Promise<Flashcard>;
+
+  /** Delete a flashcard. */
+  deleteFlashcard(id: string): Promise<void>;
+
+  /** Counts for badges (optionally course-scoped). */
+  flashcardCounts(courseId?: string | null): Promise<{ total: number; due: number }>;
 }

src/lib/db/repo.web.ts

@@ -1,29 +1,70 @@
 // Web storage implementation backed by Dexie (IndexedDB).
 //
-// We keep ONE object store ('transcripts') keyed by id. Each record holds the
+// Schema (v2):
-// full Transcript plus a derived, lowercased 'searchText' (title + all segment
+//   transcripts: full Transcript + derived lowercase 'searchText' (title + all
-// text) used for case-insensitive substring search. Search is done in-memory
+//                segment text), keyed by id, indexed by createdAt + courseId.
-// (filter over all rows) because IndexedDB has no native substring index;
+//   courses:     stored Course rows, indexed by name.
-// transcript counts are small enough on-device that this is fine.
+//   media:       persisted source audio { transcriptId, blob, mime }.
+//
+// Search is done in-memory (filter over all rows) because IndexedDB has no
+// native substring index; transcript counts are small enough on-device that
+// this is fine. searchText / blobs are never returned to callers.
 
 import Dexie from 'dexie';
 import type { Segment } from '../types';
 import { newId } from '../ids';
 import {
   parseDraft,
+  parseCourseDraft,
+  parseFlashcardDraft,
   zSegment,
   type TranscriptDraft,
   type TranscriptMeta,
   type Transcript,
+  type TranscriptPatch,
+  type Course,
+  type CourseDraft,
+  type Flashcard,
+  type FlashcardDraft,
+  type SrsState,
 } from './schema';
-import type { StorageRepo } from './repo';
+import type {
+  StorageRepo,
+  MediaInput,
+  SegmentVector,
+  VectorHit,
+} from './repo';
 
-// The shape persisted in IndexedDB: a full Transcript plus the derived
+// The shape persisted in the 'transcripts' store: a full Transcript plus the
-// searchText column. searchText is never returned to callers.
+// derived searchText column. searchText is never returned to callers.
 interface StoredTranscript extends Transcript {
   searchText: string;
 }
 
+// The shape persisted in the 'media' store. The blob/mime never leak to
+// callers; only a hasMedia flag on the transcript + an object URL are exposed.
+interface StoredMedia {
+  transcriptId: string;
+  blob: Blob;
+  mime: string;
+}
+
+// One row in the 'segvecs' store: a single segment's embedding plus the
+// denormalized courseId of its owning transcript (so course-scoped search can
+// filter without joining back to transcripts) and the embedding model tag.
+// The compound key [transcriptId+segmentId] keeps one row per segment; the
+// transcriptId / courseId / model indexes drive the cascade + filter paths.
+interface StoredSegVec {
+  transcriptId: string;
+  segmentId: string;
+  start: number;
+  end: number;
+  courseId: string | null;
+  text: string;
+  vector: Float32Array;
+  model: string;
+}
+
 // ---------------------------------------------------------------------------
 // Pure derivation helpers (shared by create/update within this file)
 // ---------------------------------------------------------------------------
@@ -35,6 +76,11 @@ function deriveSearchText(title: string, segments: Segment[]): string {
   return [title, ...segments.map((s) => s.text)].join('\n').toLowerCase();
 }
 
+/** Assign a stable id to every segment that lacks one (mutating a fresh copy). */
+function withSegmentIds(segments: Segment[]): Segment[] {
+  return segments.map((s) => (s.id ? s : { ...s, id: newId('s_') }));
+}
+
 /** Strip the derived searchText, returning the public Transcript view. */
 function toTranscript(row: StoredTranscript): Transcript {
   const { searchText: _searchText, ...rest } = row;
@@ -47,21 +93,81 @@ function toMeta(row: StoredTranscript): TranscriptMeta {
   return meta;
 }
 
+/**
+ * Cosine similarity between two vectors. Both the query and stored vectors are
+ * unit-normalized by contract, so this is just their dot product; we do not
+ * re-normalize here. Mismatched lengths fall back to the common prefix so a
+ * stale-dim row can never throw (it will simply score poorly).
+ */
+function dot(a: Float32Array, b: Float32Array): number {
+  const n = Math.min(a.length, b.length);
+  let sum = 0;
+  for (let i = 0; i < n; i++) {
+    sum += a[i]! * b[i]!;
+  }
+  return sum;
+}
+
 // ---------------------------------------------------------------------------
 // Dexie database
 // ---------------------------------------------------------------------------
 
 class WispDexie extends Dexie {
-  // Only 'id' and 'createdAt' need to be indexed: id is the primary key,
-  // createdAt powers the newest-first ordering. searchText is matched by an
-  // in-memory filter, so it does not need an index.
   transcripts!: Dexie.Table<StoredTranscript, string>;
+  courses!: Dexie.Table<Course, string>;
+  media!: Dexie.Table<StoredMedia, string>;
+  // Keyed by the compound [transcriptId+segmentId]; secondary indexes on
+  // transcriptId (cascade), courseId (scoped search) and model (backfill).
+  segvecs!: Dexie.Table<StoredSegVec, [string, string]>;
+  // Phase 3 flashcards, keyed by id; secondary indexes on transcriptId
+  // (cascade on remove) and courseId (scoping + cascade on deleteCourse).
+  flashcards!: Dexie.Table<Flashcard, string>;
 
   constructor() {
     super('wisp');
+    // v1: original single store. Kept so an existing on-disk v1 DB upgrades
+    // cleanly rather than being recreated (which would lose data).
     this.version(1).stores({
       transcripts: 'id, createdAt',
     });
+    // v2: add courseId index + courses/media stores, and backfill segment ids.
+    this.version(2)
+      .stores({
+        transcripts: 'id, createdAt, courseId',
+        courses: 'id, name',
+        media: 'transcriptId',
+      })
+      .upgrade(async (tx) => {
+        // Backfill: every pre-existing transcript gets stable ids on any
+        // segment that lacks one. courseId stays undefined => "Unsorted".
+        // No other data is touched, so existing transcripts keep everything.
+        await tx
+          .table<StoredTranscript, string>('transcripts')
+          .toCollection()
+          .modify((row) => {
+            row.segments = withSegmentIds(row.segments);
+          });
+      });
+    // v3: add the segvecs store for Phase 1 on-device semantic search. The
+    // existing transcripts/courses/media stores are re-declared unchanged so
+    // Dexie keeps them; only the new store is added. No data backfill is needed
+    // — vectors are produced lazily by the embedding pipeline.
+    this.version(3).stores({
+      transcripts: 'id, createdAt, courseId',
+      courses: 'id, name',
+      media: 'transcriptId',
+      segvecs: '[transcriptId+segmentId], transcriptId, courseId, model',
+    });
+    // v4: add the flashcards store for Phase 3 spaced repetition. Existing
+    // stores are re-declared unchanged so Dexie keeps them; only the new store
+    // is added. No backfill — flashcards are produced lazily by the study UI.
+    this.version(4).stores({
+      transcripts: 'id, createdAt, courseId',
+      courses: 'id, name',
+      media: 'transcriptId',
+      segvecs: '[transcriptId+segmentId], transcriptId, courseId, model',
+      flashcards: 'id, transcriptId, courseId',
+    });
   }
 }
 
@@ -72,6 +178,7 @@ const db = new WispDexie();
 // ---------------------------------------------------------------------------
 
 export const repo: StorageRepo = {
+  // --- transcripts ---------------------------------------------------------
   async list(): Promise<TranscriptMeta[]> {
     // Sort by createdAt then reverse for newest-first.
     const rows = await db.transcripts.orderBy('createdAt').reverse().toArray();
@@ -87,6 +194,8 @@ export const repo: StorageRepo = {
     // Validate+throw before doing anything else.
     const valid = parseDraft(draft);
     const now = Date.now();
+    // Assign a stable id to every segment that lacks one before storing.
+    const segments = withSegmentIds(valid.segments);
     const stored: StoredTranscript = {
       id: newId('t_'),
       title: valid.title,
@@ -96,28 +205,36 @@ export const repo: StorageRepo = {
       modelId: valid.modelId,
       // Preserve optional language only when present (avoid storing undefined).
       ...(valid.language !== undefined ? { language: valid.language } : {}),
-      segmentCount: valid.segments.length,
+      segmentCount: segments.length,
-      segments: valid.segments,
+      segments,
-      searchText: deriveSearchText(valid.title, valid.segments),
+      // Course/lecture metadata. courseId defaults to null ("Unsorted").
+      courseId: valid.courseId ?? null,
+      ...(valid.lectureDate !== undefined ? { lectureDate: valid.lectureDate } : {}),
+      ...(valid.instructor !== undefined ? { instructor: valid.instructor } : {}),
+      ...(valid.location !== undefined ? { location: valid.location } : {}),
+      ...(valid.tags !== undefined ? { tags: valid.tags } : {}),
+      ...(valid.lectureNumber !== undefined
+        ? { lectureNumber: valid.lectureNumber }
+        : {}),
+      hasMedia: false,
+      searchText: deriveSearchText(valid.title, segments),
     };
     await db.transcripts.put(stored);
     return toTranscript(stored);
   },
 
-  async update(
+  async update(id: string, patch: TranscriptPatch): Promise<Transcript> {
-    id: string,
-    patch: Partial<Pick<Transcript, 'title' | 'segments'>>,
-  ): Promise<Transcript> {
     const existing = await db.transcripts.get(id);
     if (!existing) {
       throw new Error(`transcript not found: ${id}`);
     }
 
     // Re-validate segments on every write so an update can never persist
-    // non-finite times that a create() would have rejected.
+    // non-finite times that a create() would have rejected, and ensure every
+    // segment carries a stable id.
     const nextSegments =
       patch.segments !== undefined
-        ? patch.segments.map((s) => zSegment.parse(s))
+        ? withSegmentIds(patch.segments.map((s) => zSegment.parse(s)))
         : existing.segments;
     const nextTitle = patch.title !== undefined ? patch.title : existing.title;
 
@@ -125,17 +242,44 @@ export const repo: StorageRepo = {
       ...existing,
       title: nextTitle,
       segments: nextSegments,
+      // Apply only the provided course/lecture meta fields.
+      ...(patch.courseId !== undefined ? { courseId: patch.courseId } : {}),
+      ...(patch.lectureDate !== undefined ? { lectureDate: patch.lectureDate } : {}),
+      ...(patch.instructor !== undefined ? { instructor: patch.instructor } : {}),
+      ...(patch.location !== undefined ? { location: patch.location } : {}),
+      ...(patch.tags !== undefined ? { tags: patch.tags } : {}),
+      ...(patch.lectureNumber !== undefined
+        ? { lectureNumber: patch.lectureNumber }
+        : {}),
       // Re-derive denormalized fields so search() and list() stay correct.
       segmentCount: nextSegments.length,
       searchText: deriveSearchText(nextTitle, nextSegments),
       updatedAt: Date.now(),
     };
     await db.transcripts.put(updated);
+    // Segments were patched => any stored vectors are now stale (they were
+    // embedded from the old text/timing). Drop them so the backfill re-embeds.
+    if (patch.segments !== undefined) {
+      await db.segvecs.where('transcriptId').equals(id).delete();
+    }
     return toTranscript(updated);
   },
 
   async remove(id: string): Promise<void> {
-    await db.transcripts.delete(id);
+    // Also delete any persisted media + vectors + flashcards for this transcript.
+    await db.transaction(
+      'rw',
+      db.transcripts,
+      db.media,
+      db.segvecs,
+      db.flashcards,
+      async () => {
+        await db.transcripts.delete(id);
+        await db.media.delete(id);
+        await db.segvecs.where('transcriptId').equals(id).delete();
+        await db.flashcards.where('transcriptId').equals(id).delete();
+      },
+    );
   },
 
   async search(query: string): Promise<TranscriptMeta[]> {
@@ -147,4 +291,319 @@ export const repo: StorageRepo = {
       : rows;
     return matched.map(toMeta);
   },
+
+  async listByCourse(courseId: string | null): Promise<TranscriptMeta[]> {
+    const rows = await db.transcripts.orderBy('createdAt').reverse().toArray();
+    const matched = rows.filter((r) =>
+      courseId === null
+        ? // "Unsorted": courseId null or undefined.
+          r.courseId == null
+        : r.courseId === courseId,
+    );
+    return matched.map(toMeta);
+  },
+
+  async reassign(transcriptId: string, courseId: string | null): Promise<void> {
+    const existing = await db.transcripts.get(transcriptId);
+    if (!existing) return;
+    await db.transaction('rw', db.transcripts, db.segvecs, async () => {
+      await db.transcripts.put({
+        ...existing,
+        courseId,
+        updatedAt: Date.now(),
+      });
+      // Keep the denormalized courseId on every vector row in sync so
+      // course-scoped search keeps matching after a move.
+      await db.segvecs
+        .where('transcriptId')
+        .equals(transcriptId)
+        .modify({ courseId });
+    });
+  },
+
+  // --- courses -------------------------------------------------------------
+  async listCourses(): Promise<Course[]> {
+    // Alphabetical by name (case-insensitive for stable UX).
+    const rows = await db.courses.toArray();
+    return rows.sort((a, b) =>
+      a.name.localeCompare(b.name, undefined, { sensitivity: 'base' }),
+    );
+  },
+
+  async createCourse(draft: CourseDraft): Promise<Course> {
+    const valid = parseCourseDraft(draft);
+    const now = Date.now();
+    const course: Course = {
+      id: newId('c_'),
+      name: valid.name,
+      ...(valid.code !== undefined ? { code: valid.code } : {}),
+      ...(valid.instructor !== undefined ? { instructor: valid.instructor } : {}),
+      ...(valid.term !== undefined ? { term: valid.term } : {}),
+      ...(valid.color !== undefined ? { color: valid.color } : {}),
+      createdAt: now,
+      updatedAt: now,
+    };
+    await db.courses.put(course);
+    return course;
+  },
+
+  async updateCourse(id: string, patch: Partial<CourseDraft>): Promise<Course> {
+    const existing = await db.courses.get(id);
+    if (!existing) {
+      throw new Error(`course not found: ${id}`);
+    }
+    const updated: Course = {
+      ...existing,
+      // Apply only the provided fields.
+      ...(patch.name !== undefined ? { name: patch.name } : {}),
+      ...(patch.code !== undefined ? { code: patch.code } : {}),
+      ...(patch.instructor !== undefined ? { instructor: patch.instructor } : {}),
+      ...(patch.term !== undefined ? { term: patch.term } : {}),
+      ...(patch.color !== undefined ? { color: patch.color } : {}),
+      updatedAt: Date.now(),
+    };
+    await db.courses.put(updated);
+    return updated;
+  },
+
+  async deleteCourse(id: string): Promise<void> {
+    // First reassign this course's transcripts to "Unsorted" (courseId=null),
+    // THEN delete the course row, so no transcript is ever orphaned.
+    await db.transaction(
+      'rw',
+      db.transcripts,
+      db.courses,
+      db.segvecs,
+      db.flashcards,
+      async () => {
+        const now = Date.now();
+        const owned = await db.transcripts
+          .filter((r) => r.courseId === id)
+          .toArray();
+        await Promise.all(
+          owned.map((r) =>
+            db.transcripts.put({ ...r, courseId: null, updatedAt: now }),
+          ),
+        );
+        // Keep denormalized vector rows consistent with their now-Unsorted owners.
+        await db.segvecs.where('courseId').equals(id).modify({ courseId: null });
+        // Flashcards in this course follow their transcripts to Unsorted.
+        await db.flashcards.where('courseId').equals(id).modify({ courseId: null });
+        await db.courses.delete(id);
+      },
+    );
+  },
+
+  // --- media (persisted source audio) -------------------------------------
+  async putMedia(transcriptId: string, input: MediaInput): Promise<void> {
+    // Web import path: bytes arrive in input.data. Wrap them in a Blob tagged
+    // with the declared mime so getMediaUrl can hand back an object URL.
+    const blob = new Blob(input.data !== undefined ? [input.data] : [], {
+      type: input.mime,
+    });
+    await db.transaction('rw', db.transcripts, db.media, async () => {
+      await db.media.put({ transcriptId, blob, mime: input.mime });
+      const existing = await db.transcripts.get(transcriptId);
+      if (existing) {
+        await db.transcripts.put({ ...existing, hasMedia: true });
+      }
+    });
+  },
+
+  async getMediaUrl(transcriptId: string): Promise<string | undefined> {
+    const row = await db.media.get(transcriptId);
+    if (!row) return undefined;
+    return URL.createObjectURL(row.blob);
+  },
+
+  async removeMedia(transcriptId: string): Promise<void> {
+    await db.transaction('rw', db.transcripts, db.media, async () => {
+      await db.media.delete(transcriptId);
+      const existing = await db.transcripts.get(transcriptId);
+      if (existing) {
+        await db.transcripts.put({ ...existing, hasMedia: false });
+      }
+    });
+  },
+
+  // --- semantic search vectors (Phase 1) ----------------------------------
+  async upsertVectors(
+    transcriptId: string,
+    model: string,
+    vectors: SegmentVector[],
+  ): Promise<void> {
+    await db.transaction('rw', db.transcripts, db.segvecs, async () => {
+      // Denormalize the transcript's current courseId onto each row so
+      // course-scoped search never has to join back to transcripts.
+      const owner = await db.transcripts.get(transcriptId);
+      const courseId = owner ? owner.courseId ?? null : null;
+      // Replace, not merge: drop every existing row for this transcript first
+      // so a re-embed with fewer segments can't leave stale rows behind.
+      await db.segvecs.where('transcriptId').equals(transcriptId).delete();
+      const rows: StoredSegVec[] = vectors.map((v) => ({
+        transcriptId,
+        segmentId: v.segmentId,
+        start: v.start,
+        end: v.end,
+        courseId,
+        text: v.text,
+        vector: v.vector,
+        model,
+      }));
+      await db.segvecs.bulkPut(rows);
+    });
+  },
+
+  async searchVectors(
+    query: Float32Array,
+    opts?: { courseId?: string | null; limit?: number },
+  ): Promise<VectorHit[]> {
+    // Scope: when opts.courseId is provided we filter (null => Unsorted, i.e.
+    // courseId == null); when it is omitted entirely we search every course.
+    let rows: StoredSegVec[];
+    if (opts && opts.courseId !== undefined) {
+      const scope = opts.courseId;
+      if (scope === null) {
+        // IndexedDB cannot index null keys, so Unsorted is filtered in memory.
+        const all = await db.segvecs.toArray();
+        rows = all.filter((r) => r.courseId === null);
+      } else {
+        rows = await db.segvecs.where('courseId').equals(scope).toArray();
+      }
+    } else {
+      rows = await db.segvecs.toArray();
+    }
+
+    const limit = opts?.limit ?? 30;
+    const hits: VectorHit[] = rows.map((r) => ({
+      transcriptId: r.transcriptId,
+      segmentId: r.segmentId,
+      start: r.start,
+      end: r.end,
+      courseId: r.courseId,
+      text: r.text,
+      score: dot(query, r.vector),
+    }));
+    hits.sort((a, b) => b.score - a.score);
+    return hits.slice(0, limit);
+  },
+
+  async clearVectors(transcriptId: string): Promise<void> {
+    await db.segvecs.where('transcriptId').equals(transcriptId).delete();
+  },
+
+  async unembeddedIds(model: string): Promise<string[]> {
+    // Transcript ids with ZERO segvecs rows for this model. We collect the set
+    // of embedded ids for the model, then return every transcript id not in it.
+    const embedded = new Set<string>();
+    await db.segvecs
+      .where('model')
+      .equals(model)
+      .each((row) => {
+        embedded.add(row.transcriptId);
+      });
+    const ids = await db.transcripts.toCollection().primaryKeys();
+    return ids.filter((id) => !embedded.has(id));
+  },
+
+  // --- flashcards + spaced repetition (Phase 3) ---------------------------
+  async createFlashcards(drafts: FlashcardDraft[]): Promise<Flashcard[]> {
+    const now = Date.now();
+    const cards: Flashcard[] = drafts.map((draft) => {
+      const valid = parseFlashcardDraft(draft);
+      // Initial SM-2 state for a brand-new card: due immediately (now). We
+      // compute the SRS inline here rather than importing learn/srs to keep the
+      // repo storage-only (no dependency on the study logic).
+      const srs: SrsState = {
+        ease: 2.5,
+        intervalDays: 0,
+        reps: 0,
+        lapses: 0,
+        due: now,
+      };
+      return {
+        id: newId('f_'),
+        transcriptId: valid.transcriptId,
+        // courseId defaults to null ("Unsorted") when absent.
+        courseId: valid.courseId ?? null,
+        ...(valid.segmentId !== undefined ? { segmentId: valid.segmentId } : {}),
+        ...(valid.start !== undefined ? { start: valid.start } : {}),
+        front: valid.front,
+        back: valid.back,
+        createdAt: now,
+        srs,
+      };
+    });
+    await db.flashcards.bulkPut(cards);
+    return cards;
+  },
+
+  async listFlashcards(opts?: {
+    transcriptId?: string;
+    courseId?: string | null;
+  }): Promise<Flashcard[]> {
+    let rows = await db.flashcards.toArray();
+    if (opts?.transcriptId !== undefined) {
+      rows = rows.filter((c) => c.transcriptId === opts.transcriptId);
+    }
+    if (opts && opts.courseId !== undefined) {
+      const scope = opts.courseId;
+      rows = rows.filter((c) =>
+        // null scope => Unsorted (courseId null or undefined).
+        scope === null ? c.courseId == null : c.courseId === scope,
+      );
+    }
+    return rows;
+  },
+
+  async listDueFlashcards(opts?: {
+    courseId?: string | null;
+    now?: number;
+    limit?: number;
+  }): Promise<Flashcard[]> {
+    const now = opts?.now ?? Date.now();
+    let rows = await db.flashcards.toArray();
+    if (opts && opts.courseId !== undefined) {
+      const scope = opts.courseId;
+      rows = rows.filter((c) =>
+        scope === null ? c.courseId == null : c.courseId === scope,
+      );
+    }
+    // Due means the next-review time is now or in the past.
+    rows = rows.filter((c) => c.srs.due <= now);
+    // Soonest-due first.
+    rows.sort((a, b) => a.srs.due - b.srs.due);
+    if (opts?.limit !== undefined) {
+      rows = rows.slice(0, opts.limit);
+    }
+    return rows;
+  },
+
+  async updateFlashcardSrs(id: string, srs: SrsState): Promise<Flashcard> {
+    const card = await db.flashcards.get(id);
+    if (!card) {
+      throw new Error(`flashcard not found: ${id}`);
+    }
+    const updated: Flashcard = { ...card, srs };
+    await db.flashcards.put(updated);
+    return updated;
+  },
+
+  async deleteFlashcard(id: string): Promise<void> {
+    await db.flashcards.delete(id);
+  },
+
+  async flashcardCounts(
+    courseId?: string | null,
+  ): Promise<{ total: number; due: number }> {
+    const now = Date.now();
+    let rows = await db.flashcards.toArray();
+    if (courseId !== undefined) {
+      rows = rows.filter((c) =>
+        courseId === null ? c.courseId == null : c.courseId === courseId,
+      );
+    }
+    const due = rows.filter((c) => c.srs.due <= now).length;
+    return { total: rows.length, due };
+  },
 };

src/lib/db/schema.ts

@@ -37,6 +37,8 @@ const zFiniteNonNeg = zFinite.refine((n) => n >= 0, {
  *  - confidence: optional, finite and within [0, 1]
  */
 export const zSegment = z.object({
+  // Stable id assigned by the repo on write (engine output omits it).
+  id: z.string().optional(),
   start: zFiniteNonNeg,
   end: zFiniteNonNeg,
   text: z.string(),
@@ -80,6 +82,15 @@ export const zTranscriptDraft = z.object({
   modelId: zModelId,
   language: z.string().optional(),
   segments: z.array(zSegment),
+  // --- Phase 0 course/lecture metadata (all optional) ---
+  /** Owning course id, or null for "Unsorted". */
+  courseId: z.string().nullable().optional(),
+  /** When the lecture happened, ms since epoch. */
+  lectureDate: zFiniteNonNeg.optional(),
+  instructor: z.string().optional(),
+  location: z.string().optional(),
+  tags: z.array(z.string()).optional(),
+  lectureNumber: zFiniteNonNeg.optional(),
 });
 
 /** Inferred TS type for a validated draft. */
@@ -93,6 +104,80 @@ export function parseDraft(input: unknown): TranscriptDraft {
   return zTranscriptDraft.parse(input);
 }
 
+// ---------------------------------------------------------------------------
+// Course (Phase 0)
+// ---------------------------------------------------------------------------
+
+/** Unsaved course shape handed to createCourse(). */
+export const zCourseDraft = z.object({
+  name: z.string().min(1),
+  code: z.string().optional(),
+  instructor: z.string().optional(),
+  term: z.string().optional(),
+  /** UI accent color (hex), optional. */
+  color: z.string().optional(),
+});
+export type CourseDraft = z.infer<typeof zCourseDraft>;
+
+/** A stored course. */
+export interface Course {
+  id: string;
+  name: string;
+  code?: string;
+  instructor?: string;
+  term?: string;
+  color?: string;
+  createdAt: number;
+  updatedAt: number;
+}
+
+/** Validate an unknown input as a CourseDraft, throwing on violation. */
+export function parseCourseDraft(input: unknown): CourseDraft {
+  return zCourseDraft.parse(input);
+}
+
+// ---------------------------------------------------------------------------
+// Flashcards + spaced repetition (Phase 3)
+// ---------------------------------------------------------------------------
+
+/** SM-2 spaced-repetition state for a card. */
+export interface SrsState {
+  /** SM-2 ease factor (starts 2.5, floored at 1.3). */
+  ease: number;
+  /** Current interval in days. */
+  intervalDays: number;
+  /** Successful repetitions in a row. */
+  reps: number;
+  /** Times the card was forgotten. */
+  lapses: number;
+  /** Next-due time, ms since epoch. */
+  due: number;
+  /** Last review time, ms since epoch. */
+  lastReviewed?: number;
+}
+
+export const zFlashcardDraft = z.object({
+  transcriptId: z.string(),
+  courseId: z.string().nullable().optional(),
+  /** Source segment for click-to-seek back to the lecture. */
+  segmentId: z.string().optional(),
+  start: zFiniteNonNeg.optional(),
+  front: z.string().min(1),
+  back: z.string().min(1),
+});
+export type FlashcardDraft = z.infer<typeof zFlashcardDraft>;
+
+/** A stored flashcard with its SRS schedule. */
+export interface Flashcard extends FlashcardDraft {
+  id: string;
+  createdAt: number;
+  srs: SrsState;
+}
+
+export function parseFlashcardDraft(input: unknown): FlashcardDraft {
+  return zFlashcardDraft.parse(input);
+}
+
 // ---------------------------------------------------------------------------
 // Stored types (output of the repo)
 // ---------------------------------------------------------------------------
@@ -114,9 +199,35 @@ export interface TranscriptMeta {
   language?: string;
   /** Number of segments in the body (denormalized for list views). */
   segmentCount: number;
+  // --- Phase 0 course/lecture metadata ---
+  /** Owning course id; null/undefined => "Unsorted". */
+  courseId?: string | null;
+  /** Lecture date, ms since epoch. */
+  lectureDate?: number;
+  instructor?: string;
+  location?: string;
+  tags?: string[];
+  lectureNumber?: number;
+  /** True when source audio has been persisted (so it can be replayed). */
+  hasMedia?: boolean;
 }
 
 /** A full stored transcript: metadata plus the segment body. */
 export interface Transcript extends TranscriptMeta {
   segments: Segment[];
 }
+
+/** The mutable fields update() accepts. */
+export type TranscriptPatch = Partial<
+  Pick<
+    Transcript,
+    | 'title'
+    | 'segments'
+    | 'courseId'
+    | 'lectureDate'
+    | 'instructor'
+    | 'location'
+    | 'tags'
+    | 'lectureNumber'
+  >
+>;

src/lib/download.native.ts

@@ -0,0 +1,33 @@
+// Native (iOS/Android) implementation of downloadText: there's no browser
+// download, so we write the content to a cache file and hand it to the OS share
+// sheet (Save to Files, AirDrop, send to another app, …). Previously this path
+// returned false and the export buttons (transcript, ICS, Anki CSV, BibTeX/RIS)
+// silently did nothing on device.
+
+import * as Sharing from 'expo-sharing';
+import { writeAsStringAsync, cacheDirectory } from 'expo-file-system/legacy';
+
+/** Sanitize a filename for use as a path segment (keep word chars, dot, dash). */
+function safeName(name: string): string {
+  const cleaned = name.replace(/[^\w.\-]+/g, '_').replace(/^_+|_+$/g, '');
+  return cleaned.length > 0 ? cleaned : 'export';
+}
+
+export async function downloadText(
+  filename: string,
+  mime: string,
+  content: string,
+): Promise<boolean> {
+  try {
+    const dir = cacheDirectory;
+    if (!dir || !(await Sharing.isAvailableAsync())) return false;
+    const uri = `${dir}${safeName(filename)}`;
+    await writeAsStringAsync(uri, content);
+    await Sharing.shareAsync(uri, { mimeType: mime, dialogTitle: filename });
+    return true;
+  } catch {
+    // User-cancelled share or any IO error: report failure so callers can decide
+    // (today they ignore it, but the return keeps the contract honest).
+    return false;
+  }
+}

src/lib/download.ts

@@ -1,11 +1,12 @@
-import { Platform } from 'react-native';
+// Web implementation (also the tsc base; native uses download.native.ts).
-
+//
-/**
+// Triggers a client-side file download on web (no server, nothing uploaded).
- * Trigger a client-side file download on web (no server, nothing uploaded).
+// Async so the signature matches the native share-sheet implementation.
- * Returns false on native, where the caller should fall back to share/save.
+export async function downloadText(
- */
+  filename: string,
-export function downloadText(filename: string, mime: string, content: string): boolean {
+  mime: string,
-  if (Platform.OS !== 'web') return false;
+  content: string,
+): Promise<boolean> {
   const blob = new Blob([content], { type: mime });
   const url = URL.createObjectURL(blob);
   const a = document.createElement('a');

src/lib/embedding/engine.ts

@@ -0,0 +1,28 @@
+// The embedding-engine interface: hides the platform-specific sentence-embedding
+// backend (transformers.js feature-extraction on web; a follow-up native impl)
+// behind one contract, mirroring TranscriptionEngine. Used for on-device
+// semantic search (ROADMAP Phase 1) — vectors never leave the device.
+
+export interface EmbeddingEngine {
+  readonly platform: 'web' | 'native';
+  /** Vector dimensionality (must match EMBED_DIM). */
+  readonly dim: number;
+
+  /** Synchronous check: is the model loaded in memory? */
+  isLoaded(): boolean;
+
+  /** Ensure the model is loaded. `onProgress` (0..1) fires during download. */
+  loadModel(onProgress?: (p: number) => void): Promise<void>;
+
+  /**
+   * Embed texts into unit-normalized, mean-pooled vectors. `kind` adds the
+   * asymmetric prefix the model expects (e5: "query:" for search text,
+   * "passage:" for stored segments) so retrieval is well-calibrated.
+   */
+  embed(texts: string[], kind: 'query' | 'passage'): Promise<Float32Array[]>;
+}
+
+/** Embedding model id — also stored alongside vectors to detect model changes. */
+export const EMBED_MODEL = 'Xenova/multilingual-e5-small';
+/** Output dimensionality of EMBED_MODEL. */
+export const EMBED_DIM = 384;

src/lib/embedding/engineImpl.native.ts

@@ -0,0 +1,28 @@
+// NATIVE-ONLY embedding engine stub. On-device sentence embeddings on native
+// (a transformers/onnxruntime-react-native or whisper.rn-style binding) are a
+// follow-up; until then this is a type-correct placeholder that throws clearly
+// so the rest of the app stays platform-agnostic and never silently mis-embeds.
+//
+// Selected by Metro at build time for native targets; never imported by tests.
+
+import type { EmbeddingEngine } from './engine';
+import { EMBED_DIM } from './engine';
+
+const NOT_AVAILABLE = 'On-device embeddings are not available on native yet';
+
+export const engine: EmbeddingEngine = {
+  platform: 'native',
+  dim: EMBED_DIM,
+
+  isLoaded(): boolean {
+    return false;
+  },
+
+  async loadModel(): Promise<void> {
+    throw new Error(NOT_AVAILABLE);
+  },
+
+  async embed(): Promise<Float32Array[]> {
+    throw new Error(NOT_AVAILABLE);
+  },
+};

src/lib/embedding/engineImpl.ts

@@ -0,0 +1,3 @@
+// Base resolver for TypeScript. Metro picks engineImpl.web.ts / engineImpl.native.ts
+// by platform extension at build time; tsc resolves this file (defaults to web).
+export { engine } from './engineImpl.web';

src/lib/embedding/engineImpl.web.ts

@@ -0,0 +1,114 @@
+// WEB-ONLY embedding engine, backed by transformers.js (@huggingface/
+// transformers) running the multilingual-e5-small feature-extraction model in
+// the browser (WebGPU when available, else WASM). Produces unit-normalized,
+// mean-pooled 384-d vectors for on-device semantic search (Phase 1).
+//
+// WHY WE LOAD IT FROM A CDN AT RUNTIME (not a static import):
+// transformers.js depends on onnxruntime-web, which uses a *computed* dynamic
+// import (`import(/*webpackIgnore*/ a)`) and ships WASM — Metro (Expo's web
+// bundler) cannot statically bundle either and fails the build. So we never let
+// Metro see the package: we load the ESM build from a CDN at runtime via a
+// dynamic import hidden behind `new Function` (so Metro's static analyzer can't
+// trip over it). This mirrors src/lib/transcription/engineImpl.web.ts exactly.
+//
+// NOTE: the page must be cross-origin isolated (COOP + COEP) for multi-threaded
+// WASM; see docker/nginx.conf. This module is web-only and is NEVER imported by
+// any vitest test.
+
+import type { EmbeddingEngine } from './engine';
+import { EMBED_DIM, EMBED_MODEL } from './engine';
+
+// Pin the transformers.js version we load at runtime (matches the ASR engine).
+const TRANSFORMERS_CDN = 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.2.0';
+
+// `new Function` hides the dynamic import() specifier from Metro's bundler so it
+// never tries to resolve/transform transformers.js or onnxruntime-web.
+const runtimeImport = new Function('u', 'return import(u)') as (u: string) => Promise<TransformersModule>;
+
+// Minimal structural types for the bits of transformers.js we use.
+interface FeatureTensor {
+  /** Nested arrays: one row of floats per input text. */
+  tolist(): number[][];
+}
+type FeatureExtractor = (
+  texts: string[],
+  opts: { pooling: 'mean'; normalize: boolean },
+) => Promise<FeatureTensor>;
+interface PipelineOptions {
+  device?: string;
+  dtype?: string;
+  progress_callback?: (e: { status?: string; progress?: number }) => void;
+}
+interface TransformersModule {
+  pipeline: (task: string, model: string, opts?: PipelineOptions) => Promise<FeatureExtractor>;
+  env: { allowLocalModels: boolean };
+}
+
+let libPromise: Promise<TransformersModule> | null = null;
+async function lib(): Promise<TransformersModule> {
+  if (!libPromise) {
+    libPromise = runtimeImport(TRANSFORMERS_CDN).then((m) => {
+      // Never read models off the local filesystem in the browser.
+      m.env.allowLocalModels = false;
+      return m;
+    });
+  }
+  return libPromise;
+}
+
+/** The single cached feature-extraction pipeline (one model). */
+let extractor: FeatureExtractor | null = null;
+let cachedWebGpu: boolean | undefined;
+
+async function detectWebGpu(): Promise<boolean> {
+  if (cachedWebGpu !== undefined) return cachedWebGpu;
+  try {
+    if (typeof navigator === 'undefined' || !('gpu' in navigator)) {
+      cachedWebGpu = false;
+      return false;
+    }
+    const gpu = (navigator as { gpu?: { requestAdapter(): Promise<unknown> } }).gpu;
+    const adapter = await gpu?.requestAdapter();
+    cachedWebGpu = adapter != null;
+  } catch {
+    cachedWebGpu = false;
+  }
+  return cachedWebGpu;
+}
+
+export const engine: EmbeddingEngine = {
+  platform: 'web',
+  dim: EMBED_DIM,
+
+  isLoaded(): boolean {
+    return extractor != null;
+  },
+
+  async loadModel(onProgress?: (p: number) => void): Promise<void> {
+    if (extractor != null) return; // idempotent
+    const { pipeline } = await lib();
+    const webgpu = await detectWebGpu();
+    extractor = await pipeline('feature-extraction', EMBED_MODEL, {
+      // WebGPU + fp16 when available; otherwise 8-bit weights on WASM, which
+      // stays small to download and runs acceptably on a plain CPU.
+      device: webgpu ? 'webgpu' : 'wasm',
+      dtype: webgpu ? 'fp16' : 'q8',
+      progress_callback: (e) => {
+        if (e.status === 'progress' && e.progress != null) onProgress?.(e.progress / 100);
+      },
+    });
+  },
+
+  async embed(texts: string[], kind: 'query' | 'passage'): Promise<Float32Array[]> {
+    if (extractor == null) {
+      throw new Error('Embedding model is not loaded; call loadModel() first.');
+    }
+    if (texts.length === 0) return [];
+    // e5 asymmetric convention: prefix with "query: " or "passage: ".
+    const prefix = `${kind}: `;
+    const prefixed = texts.map((t) => prefix + t);
+    const tensor = await extractor(prefixed, { pooling: 'mean', normalize: true });
+    // tolist() yields one number[] row per input; convert each to Float32Array.
+    return tensor.tolist().map((row) => Float32Array.from(row));
+  },
+};

src/lib/embedding/index.ts

@@ -0,0 +1,12 @@
+// Public entry point for the embedding engine.
+//
+// `./engineImpl` is resolved by Metro to engineImpl.web.ts or
+// engineImpl.native.ts by platform extension; the base engineImpl.ts re-export
+// (web) is what TypeScript resolves for typechecking. Consumers call
+// getEmbeddingEngine() and stay platform-agnostic.
+import { engine } from './engineImpl';
+
+/** Return the platform-resolved embedding engine. */
+export const getEmbeddingEngine = () => engine;
+
+export * from './engine';

src/lib/enrich/bib.test.ts

@@ -0,0 +1,78 @@
+import { describe, it, expect } from 'vitest';
+import { toBibtex, toRis } from './bib';
+import type { RefMetadata } from './types';
+
+describe('toBibtex', () => {
+  it('returns "" for []', () => {
+    expect(toBibtex([])).toBe('');
+  });
+
+  it('emits an @article with a first-author+year key', () => {
+    const refs: RefMetadata[] = [
+      {
+        title: 'Attention Is All You Need',
+        authors: ['Ashish Vaswani', 'Noam Shazeer'],
+        year: 2017,
+        source: 'Crossref',
+        url: 'https://doi.org/10.x',
+      },
+    ];
+    const out = toBibtex(refs);
+    expect(out).toContain('@article{vaswani2017,');
+    expect(out).toContain('title = {Attention Is All You Need}');
+    expect(out).toContain('author = {Ashish Vaswani and Noam Shazeer}');
+    expect(out).toContain('year = {2017}');
+    expect(out).toContain('url = {https://doi.org/10.x}');
+  });
+
+  it('uses @misc and ref<index> key when no authors', () => {
+    const out = toBibtex([{ title: 'A book', source: 'Open Library' }]);
+    expect(out).toContain('@misc{ref0,');
+    expect(out).toContain('title = {A book}');
+  });
+
+  it('omits missing fields', () => {
+    const out = toBibtex([{ source: 'X' }]);
+    expect(out).toContain('@misc{ref0,');
+    expect(out).not.toContain('title =');
+    expect(out).not.toContain('year =');
+  });
+
+  it('brace-escapes special characters in values', () => {
+    const out = toBibtex([{ title: 'a {b} c', source: 'X' }]);
+    expect(out).toContain('title = {a \\{b\\} c}');
+  });
+});
+
+describe('toRis', () => {
+  it('returns "" for []', () => {
+    expect(toRis([])).toBe('');
+  });
+
+  it('emits a JOUR record with one AU line per author', () => {
+    const out = toRis([
+      {
+        title: 'T',
+        authors: ['A One', 'B Two'],
+        year: 2020,
+        source: 'X',
+        url: 'http://u',
+      },
+    ]);
+    expect(out).toContain('TY  - JOUR');
+    expect(out).toContain('TI  - T');
+    expect(out).toContain('AU  - A One');
+    expect(out).toContain('AU  - B Two');
+    expect(out).toContain('PY  - 2020');
+    expect(out).toContain('UR  - http://u');
+    expect(out.split('\n')).toContain('ER  - ');
+  });
+
+  it('uses GEN and omits missing fields', () => {
+    const out = toRis([{ source: 'X' }]);
+    expect(out).toContain('TY  - GEN');
+    expect(out).not.toContain('TI  -');
+    expect(out).not.toContain('AU  -');
+    expect(out).toContain('ER  - ');
+  });
+});

src/lib/enrich/bib.ts

@@ -0,0 +1,80 @@
+// Pure bibliography serialization: BibTeX and RIS. No network, no I/O.
+
+import type { RefMetadata } from './types';
+
+/** Serialize references to BibTeX. Handles missing fields and []. */
+export function toBibtex(refs: RefMetadata[]): string {
+  // BibTeX entry keys MUST be unique — same-surname/same-year refs would
+  // otherwise collide and most processors drop the duplicate on import. Track
+  // emitted keys and disambiguate collisions with an a/b/c… suffix.
+  const used = new Map<string, number>();
+  return refs
+    .map((r, i) => {
+      let key = bibKey(r, i);
+      const seen = used.get(key);
+      if (seen !== undefined) {
+        const suffix = String.fromCharCode(97 + seen); // 0 -> 'a', 1 -> 'b', …
+        used.set(key, seen + 1);
+        key = `${key}${suffix}`;
+      } else {
+        used.set(key, 0);
+      }
+      return bibEntry(r, key);
+    })
+    .join('\n\n');
+}
+
+function bibEntry(r: RefMetadata, key: string): string {
+  const type = r.authors && r.authors.length ? 'article' : 'misc';
+
+  const fields: string[] = [];
+  if (r.title) fields.push(`  title = {${brace(r.title)}}`);
+  if (r.authors && r.authors.length) {
+    fields.push(`  author = {${brace(r.authors.join(' and '))}}`);
+  }
+  if (r.year !== undefined) fields.push(`  year = {${r.year}}`);
+  if (r.url) fields.push(`  url = {${brace(r.url)}}`);
+
+  return `@${type}{${key},\n${fields.join(',\n')}\n}`;
+}
+
+function bibKey(r: RefMetadata, index: number): string {
+  const first = r.authors && r.authors[0];
+  if (first) {
+    // Last token of the first author's name, alphanumerics only.
+    const tokens = first.trim().split(/[\s,]+/);
+    const surname = (tokens[tokens.length - 1] ?? first)
+      .replace(/[^A-Za-z0-9]/g, '')
+      .toLowerCase();
+    if (surname) return `${surname}${r.year ?? ''}`;
+  }
+  return `ref${index}`;
+}
+
+/** Escape characters special to BibTeX brace-delimited values. */
+function brace(s: string): string {
+  // Inside `{...}` we still must escape a stray backslash and the braces.
+  return s.replace(/\\/g, '\\\\').replace(/[{}]/g, (m) => `\\${m}`);
+}
+
+// ---------------------------------------------------------------------------
+// RIS
+// ---------------------------------------------------------------------------
+
+/** Serialize references to RIS. Handles missing fields and []. */
+export function toRis(refs: RefMetadata[]): string {
+  return refs.map(risEntry).join('\n\n');
+}
+
+function risEntry(r: RefMetadata): string {
+  const ty = r.authors && r.authors.length ? 'JOUR' : 'GEN';
+  const lines: string[] = [`TY  - ${ty}`];
+  if (r.title) lines.push(`TI  - ${r.title}`);
+  if (r.authors) {
+    for (const a of r.authors) lines.push(`AU  - ${a}`);
+  }
+  if (r.year !== undefined) lines.push(`PY  - ${r.year}`);
+  if (r.url) lines.push(`UR  - ${r.url}`);
+  lines.push('ER  - ');
+  return lines.join('\n');
+}

src/lib/enrich/citations.test.ts

@@ -0,0 +1,84 @@
+import { describe, it, expect } from 'vitest';
+import { detectCitations } from './citations';
+import type { Segment } from '../types';
+
+const seg = (start: number, text: string, id?: string): Segment => ({
+  id,
+  start,
+  end: start + 1,
+  text,
+});
+
+describe('detectCitations', () => {
+  it('returns [] for empty input', () => {
+    expect(detectCitations([])).toEqual([]);
+  });
+
+  it('extracts a bare DOI', () => {
+    const out = detectCitations([seg(0, 'See 10.1038/nphys1170 for details.')]);
+    const doi = out.find((c) => c.kind === 'doi');
+    expect(doi?.value).toBe('10.1038/nphys1170');
+  });
+
+  it('extracts a DOI from a doi.org URL and a doi: prefix', () => {
+    const a = detectCitations([seg(0, 'https://doi.org/10.1145/3292500.3330701')]);
+    expect(a[0]!.value).toBe('10.1145/3292500.3330701');
+    const b = detectCitations([seg(0, 'doi: 10.1000/xyz123')]);
+    expect(b[0]!.value).toBe('10.1000/xyz123');
+  });
+
+  it('trims trailing punctuation from a DOI', () => {
+    const out = detectCitations([seg(0, '(see 10.1038/nphys1170).')]);
+    expect(out[0]!.value).toBe('10.1038/nphys1170');
+  });
+
+  it('extracts an explicit arXiv id', () => {
+    const out = detectCitations([seg(0, 'Read arXiv:2103.00020 today.')]);
+    const ax = out.find((c) => c.kind === 'arxiv');
+    expect(ax?.value).toBe('2103.00020');
+  });
+
+  it('extracts a bare arXiv id', () => {
+    const out = detectCitations([seg(0, 'the paper 1706.03762 introduced it')]);
+    const ax = out.find((c) => c.kind === 'arxiv');
+    expect(ax?.value).toBe('1706.03762');
+  });
+
+  it('extracts ISBN-13 and ISBN-10, normalized without hyphens', () => {
+    const a = detectCitations([seg(0, 'ISBN 978-0-13-468599-1')]);
+    expect(a.find((c) => c.kind === 'isbn')?.value).toBe('9780134685991');
+    const b = detectCitations([seg(0, 'ISBN-10: 0-262-03384-4')]);
+    expect(b.find((c) => c.kind === 'isbn')?.value).toBe('0262033844');
+  });
+
+  it('extracts author-year (parenthesized)', () => {
+    const out = detectCitations([seg(0, 'As Vaswani (2017) showed,')]);
+    const ay = out.find((c) => c.kind === 'author-year');
+    expect(ay?.value).toBe('Vaswani (2017)');
+  });
+
+  it('extracts "Author et al., Year"', () => {
+    const out = detectCitations([seg(0, 'per Smith et al., 2021 this holds')]);
+    const ay = out.find((c) => c.kind === 'author-year');
+    expect(ay?.value).toBe('Smith et al. (2021)');
+  });
+
+  it('dedupes by value', () => {
+    const out = detectCitations([
+      seg(0, '10.1038/nphys1170'),
+      seg(1, 'again 10.1038/nphys1170'),
+    ]);
+    expect(out.filter((c) => c.kind === 'doi')).toHaveLength(1);
+  });
+
+  it('tags sourceSegmentId and start', () => {
+    const out = detectCitations([seg(9, 'arXiv:2103.00020', 'segA')]);
+    expect(out[0]!.sourceSegmentId).toBe('segA');
+    expect(out[0]!.start).toBe(9);
+  });
+
+  it('does not treat plain prose years as citations', () => {
+    const out = detectCitations([seg(0, 'In 2017 the weather was nice.')]);
+    expect(out.filter((c) => c.kind === 'author-year')).toHaveLength(0);
+  });
+});

src/lib/enrich/citations.ts

@@ -0,0 +1,165 @@
+// Pure citation extraction over lecture segments. Regex-only; no network.
+// Detects DOIs, arXiv ids, ISBN-10/13, and "Author (Year)" mentions.
+
+import type { Segment } from '../types';
+import type { Citation } from './types';
+
+/** Detect and normalize citations across all segments. Deduped by value. */
+export function detectCitations(segments: Segment[]): Citation[] {
+  if (!segments.length) return [];
+
+  const out: Citation[] = [];
+  const seen = new Set<string>();
+
+  const push = (c: Citation) => {
+    const key = `${c.kind}:${c.value.toLowerCase()}`;
+    if (seen.has(key)) return;
+    seen.add(key);
+    out.push(c);
+  };
+
+  for (const seg of segments) {
+    const text = seg.text ?? '';
+    if (!text) continue;
+
+    for (const c of detectDois(text)) push(tag(c, seg));
+    for (const c of detectArxiv(text)) push(tag(c, seg));
+    for (const c of detectIsbn(text)) push(tag(c, seg));
+    for (const c of detectAuthorYear(text)) push(tag(c, seg));
+  }
+
+  return out;
+}
+
+function tag(
+  c: Omit<Citation, 'sourceSegmentId' | 'start'>,
+  seg: Segment,
+): Citation {
+  return { ...c, sourceSegmentId: seg.id, start: seg.start };
+}
+
+// ---------------------------------------------------------------------------
+// DOI
+// ---------------------------------------------------------------------------
+
+function detectDois(text: string): Omit<Citation, 'sourceSegmentId' | 'start'>[] {
+  const out: Omit<Citation, 'sourceSegmentId' | 'start'>[] = [];
+  // DOIs start with 10.<registrant>/<suffix>. Allow an optional doi: / URL prefix.
+  const re = /\b(?:doi:\s*|https?:\/\/(?:dx\.)?doi\.org\/)?(10\.\d{4,9}\/[^\s"<>]+)/gi;
+  let m: RegExpExecArray | null;
+  while ((m = re.exec(text))) {
+    // Trim common trailing punctuation that isn't part of the DOI.
+    const value = m[1]!.replace(/[.,;)\]]+$/, '');
+    out.push({ kind: 'doi', value, raw: m[0] });
+  }
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// arXiv
+// ---------------------------------------------------------------------------
+
+function detectArxiv(text: string): Omit<Citation, 'sourceSegmentId' | 'start'>[] {
+  const out: Omit<Citation, 'sourceSegmentId' | 'start'>[] = [];
+
+  // Explicit "arXiv:NNNN.NNNNN" (with optional version) — most reliable.
+  const reExplicit = /\barxiv:\s*(\d{4}\.\d{4,5}(?:v\d+)?)\b/gi;
+  let m: RegExpExecArray | null;
+  while ((m = reExplicit.exec(text))) {
+    out.push({ kind: 'arxiv', value: m[1]!, raw: m[0] });
+  }
+
+  // Bare "NNNN.NNNNN" — only when not already captured as part of an arXiv: hit.
+  const reBare = /(?<![\w.])(\d{4}\.\d{4,5}(?:v\d+)?)(?![\w.])/g;
+  while ((m = reBare.exec(text))) {
+    const value = m[1]!;
+    // Skip if this position is the tail of an explicit arXiv match.
+    const preceding = text.slice(Math.max(0, m.index - 8), m.index).toLowerCase();
+    if (preceding.includes('arxiv')) continue;
+    out.push({ kind: 'arxiv', value, raw: m[0] });
+  }
+
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// ISBN
+// ---------------------------------------------------------------------------
+
+function detectIsbn(text: string): Omit<Citation, 'sourceSegmentId' | 'start'>[] {
+  const out: Omit<Citation, 'sourceSegmentId' | 'start'>[] = [];
+  // ISBN-13 (978/979 prefix) and ISBN-10, hyphens/spaces allowed internally.
+  const re =
+    /\bISBN(?:-1[03])?:?\s*((?:97[89][-\s]?)?(?:\d[-\s]?){9}[\dXx])\b/g;
+  let m: RegExpExecArray | null;
+  while ((m = re.exec(text))) {
+    const value = m[1]!.replace(/[-\s]/g, '').toUpperCase();
+    if (value.length === 10 || value.length === 13) {
+      out.push({ kind: 'isbn', value, raw: m[0] });
+    }
+  }
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// Author-year
+// ---------------------------------------------------------------------------
+
+function detectAuthorYear(
+  text: string,
+): Omit<Citation, 'sourceSegmentId' | 'start'>[] {
+  const out: Omit<Citation, 'sourceSegmentId' | 'start'>[] = [];
+
+  // "Smith (2020)", "Smith and Jones (2019)", "Smith et al., 2021",
+  // "Smith et al. (2021)". Author = Capitalized word(s).
+  const re =
+    /\b([A-Z][a-z]+(?:[-\s][A-Z][a-z]+)*(?:\s+(?:and|&)\s+[A-Z][a-z]+)?(?:\s+et al\.?)?)\s*[,(]?\s*\(?(\b(?:18|19|20)\d{2})\)?/g;
+  let m: RegExpExecArray | null;
+  while ((m = re.exec(text))) {
+    const author = stripLeadingStopword(m[1]!.trim().replace(/\s+/g, ' '));
+    const year = m[2]!;
+    // Require either a parenthesized year or an "et al"/"and" multi-author cue
+    // to avoid matching plain prose like "World War 2 1945".
+    const cue = m[0];
+    const looksLikeCitation =
+      /\(\s*(?:18|19|20)\d{2}\s*\)/.test(cue) ||
+      /et al/i.test(cue) ||
+      /\band\b|&/.test(cue);
+    if (!looksLikeCitation) continue;
+    const value = `${author} (${year})`;
+    out.push({ kind: 'author-year', value, raw: m[0].trim() });
+  }
+  return out;
+}
+
+// Common capitalized sentence-starters that can be mistaken for a surname when
+// they immediately precede the real author (e.g. "As Vaswani (2017)").
+const LEADING_STOPWORDS = new Set([
+  'As',
+  'In',
+  'By',
+  'See',
+  'Per',
+  'From',
+  'After',
+  'Before',
+  'And',
+  'The',
+  'This',
+  'That',
+  'These',
+  'Those',
+  'Both',
+  'When',
+  'While',
+  'Since',
+]);
+
+function stripLeadingStopword(author: string): string {
+  const tokens = author.split(' ');
+  // Only strip when something real remains and the next token is itself a name.
+  if (tokens.length > 1 && tokens[0] && LEADING_STOPWORDS.has(tokens[0])) {
+    return tokens.slice(1).join(' ');
+  }
+  return author;
+}

src/lib/enrich/dates.test.ts

@@ -0,0 +1,145 @@
+import { describe, it, expect } from 'vitest';
+import { detectEvents } from './dates';
+import type { Segment } from '../types';
+
+const seg = (start: number, text: string, id?: string): Segment => ({
+  id,
+  start,
+  end: start + 1,
+  text,
+});
+
+// Anchor = 2026-06-13 (Saturday) 00:00 UTC. All assertions derive from this.
+const ANCHOR = Date.UTC(2026, 5, 13, 0, 0, 0, 0);
+
+const utc = (y: number, mo: number, d: number, h = 0, mi = 0) =>
+  Date.UTC(y, mo, d, h, mi, 0, 0);
+
+describe('detectEvents', () => {
+  it('returns [] for empty input', () => {
+    expect(detectEvents([], ANCHOR)).toEqual([]);
+  });
+
+  it('returns [] for non-finite anchor', () => {
+    expect(detectEvents([seg(0, 'exam tomorrow')], Number.NaN)).toEqual([]);
+  });
+
+  it('resolves ISO YYYY-MM-DD to UTC midnight', () => {
+    const out = detectEvents([seg(0, 'The exam is on 2026-06-20.')], ANCHOR);
+    expect(out).toHaveLength(1);
+    expect(out[0]!.dateMs).toBe(utc(2026, 5, 20));
+    expect(out[0]!.allDay).toBe(true);
+    expect(out[0]!.confidence).toBeGreaterThanOrEqual(0.9);
+  });
+
+  it('resolves "March 3" using the anchor year (no roll within 6mo)', () => {
+    const out = detectEvents([seg(0, 'deadline on March 3')], ANCHOR);
+    expect(out).toHaveLength(1);
+    // March 3 2026 is ~3mo before June 13 2026 -> stays in the anchor year.
+    expect(out[0]!.dateMs).toBe(utc(2026, 2, 3));
+  });
+
+  it('rolls a date more than 6 months before the anchor to next year', () => {
+    // Anchor late in the year: "January 5" resolves ~10mo back -> roll to next.
+    const novAnchor = Date.UTC(2026, 10, 1); // 2026-11-01
+    const out = detectEvents([seg(0, 'deadline on January 5')], novAnchor);
+    expect(out[0]!.dateMs).toBe(utc(2027, 0, 5));
+  });
+
+  it('keeps an explicit year even if far past', () => {
+    const out = detectEvents([seg(0, 'paper from March 3, 2020')], ANCHOR);
+    expect(out[0]!.dateMs).toBe(utc(2020, 2, 3));
+  });
+
+  it('resolves day-first "3 March"', () => {
+    const out = detectEvents([seg(0, 'submission due 20 July')], ANCHOR);
+    expect(out[0]!.dateMs).toBe(utc(2026, 6, 20));
+  });
+
+  it('does NOT roll a date within the next 6 months', () => {
+    const out = detectEvents([seg(0, 'exam on June 20')], ANCHOR);
+    expect(out[0]!.dateMs).toBe(utc(2026, 5, 20));
+  });
+
+  it('resolves "on the 14th" within the anchor month', () => {
+    const out = detectEvents([seg(0, 'assignment due on the 14th')], ANCHOR);
+    expect(out[0]!.dateMs).toBe(utc(2026, 5, 14));
+  });
+
+  it('resolves relative "tomorrow" from the anchor', () => {
+    const out = detectEvents([seg(0, 'quiz tomorrow')], ANCHOR);
+    expect(out[0]!.dateMs).toBe(utc(2026, 5, 14));
+    expect(out[0]!.confidence).toBeLessThan(0.9); // relative
+  });
+
+  it('resolves "today" to the anchor day', () => {
+    const out = detectEvents([seg(0, 'homework due today')], ANCHOR);
+    expect(out[0]!.dateMs).toBe(utc(2026, 5, 13));
+  });
+
+  it('resolves "next week" as +7 days', () => {
+    const out = detectEvents([seg(0, 'the exam is next week')], ANCHOR);
+    expect(out[0]!.dateMs).toBe(utc(2026, 5, 20));
+  });
+
+  it('resolves "in 3 days" and "in 2 weeks"', () => {
+    const a = detectEvents([seg(0, 'due in 3 days')], ANCHOR);
+    expect(a[0]!.dateMs).toBe(utc(2026, 5, 16));
+    const b = detectEvents([seg(0, 'due in 2 weeks')], ANCHOR);
+    expect(b[0]!.dateMs).toBe(utc(2026, 5, 27));
+  });
+
+  it('resolves "next Monday" to the Monday of the following week', () => {
+    // Anchor Sat 2026-06-13. Next Monday = 2026-06-22 (skip the 15th).
+    const out = detectEvents([seg(0, 'exam next Monday')], ANCHOR);
+    expect(out[0]!.dateMs).toBe(utc(2026, 5, 22));
+  });
+
+  it('captures a clock time and sets allDay=false', () => {
+    const out = detectEvents([seg(0, 'exam on June 20 at 3pm')], ANCHOR);
+    expect(out[0]!.allDay).toBe(false);
+    expect(out[0]!.dateMs).toBe(utc(2026, 5, 20, 15, 0));
+  });
+
+  it('parses 24h time "14:00"', () => {
+    const out = detectEvents([seg(0, 'meeting 2026-06-20 14:00')], ANCHOR);
+    expect(out[0]!.allDay).toBe(false);
+    expect(out[0]!.dateMs).toBe(utc(2026, 5, 20, 14, 0));
+  });
+
+  it('uses the surrounding clause as the title', () => {
+    const out = detectEvents(
+      [seg(0, 'Welcome everyone. The final exam is on June 20. See you.')],
+      ANCHOR,
+    );
+    expect(out[0]!.title.toLowerCase()).toContain('final exam');
+  });
+
+  it('boosts confidence when a deadline keyword is present', () => {
+    const withKw = detectEvents([seg(0, 'the exam is on June 20')], ANCHOR)[0]!;
+    const without = detectEvents([seg(0, 'we met on June 20')], ANCHOR)[0]!;
+    expect(withKw.confidence).toBeGreaterThan(without.confidence);
+  });
+
+  it('tags sourceSegmentId and start', () => {
+    const out = detectEvents([seg(42, 'exam on June 20', 'segX')], ANCHOR);
+    expect(out[0]!.sourceSegmentId).toBe('segX');
+    expect(out[0]!.start).toBe(42);
+  });
+
+  it('dedupes identical (dateMs + title)', () => {
+    const out = detectEvents(
+      [seg(0, 'exam on June 20'), seg(1, 'exam on June 20')],
+      ANCHOR,
+    );
+    expect(out).toHaveLength(1);
+  });
+
+  it('does not read the wall clock (output depends only on the anchor)', () => {
+    const a = detectEvents([seg(0, 'quiz tomorrow')], ANCHOR);
+    const other = Date.UTC(2030, 0, 1);
+    const b = detectEvents([seg(0, 'quiz tomorrow')], other);
+    expect(a[0]!.dateMs).toBe(utc(2026, 5, 14));
+    expect(b[0]!.dateMs).toBe(Date.UTC(2030, 0, 2));
+  });
+});

src/lib/enrich/dates.ts

@@ -0,0 +1,427 @@
+// Pure deadline/date detection over lecture segments. NO wall-clock reads:
+// every timestamp is derived from `anchorMs` (the lecture date) plus explicit
+// year/month/day/hour components found in the text. Constructing Date objects
+// from those explicit components is fine; calling Date.now() / new Date() with
+// no args is NOT.
+
+import type { Segment } from '../types';
+import type { CandidateEvent } from './types';
+
+const MONTHS: Record<string, number> = {
+  january: 0,
+  february: 1,
+  march: 2,
+  april: 3,
+  may: 4,
+  june: 5,
+  july: 6,
+  august: 7,
+  september: 8,
+  october: 9,
+  november: 10,
+  december: 11,
+  jan: 0,
+  feb: 1,
+  mar: 2,
+  apr: 3,
+  jun: 5,
+  jul: 6,
+  aug: 7,
+  sep: 8,
+  sept: 8,
+  oct: 9,
+  nov: 10,
+  dec: 11,
+};
+
+const WEEKDAYS: Record<string, number> = {
+  sunday: 0,
+  monday: 1,
+  tuesday: 2,
+  wednesday: 3,
+  thursday: 4,
+  friday: 5,
+  saturday: 6,
+};
+
+const DEADLINE_WORDS = [
+  'exam',
+  'deadline',
+  'due',
+  'assignment',
+  'midterm',
+  'final',
+  'quiz',
+  'submit',
+  'submission',
+  'project',
+  'homework',
+  'test',
+];
+
+const SIX_MONTHS_MS = 1000 * 60 * 60 * 24 * 30 * 6;
+
+interface TimeOfDay {
+  hour: number;
+  minute: number;
+}
+
+/** A partial match produced by a detector before titling/dating. */
+interface RawMatch {
+  /** Absolute resolved instant in ms. */
+  dateMs: number;
+  /** Whether a clock time was found (=> not all-day). */
+  hasTime: boolean;
+  /** Matched substring (for `raw`). */
+  matched: string;
+  /** Index of the match within the segment text (for clause extraction). */
+  index: number;
+  /** Base confidence before deadline-keyword boosting. */
+  baseConfidence: number;
+}
+
+/**
+ * Detect candidate dates/deadlines across all segments.
+ *
+ * @param segments  lecture segments (text + timing).
+ * @param anchorMs  the lecture date in ms; relative dates anchor here and
+ *                  absolute dates borrow this year when none is written.
+ */
+export function detectEvents(segments: Segment[], anchorMs: number): CandidateEvent[] {
+  if (!segments.length || !Number.isFinite(anchorMs)) return [];
+
+  const out: CandidateEvent[] = [];
+  const seen = new Set<string>();
+
+  for (const seg of segments) {
+    const text = seg.text ?? '';
+    if (!text) continue;
+
+    const matches = [
+      ...detectIso(text, anchorMs),
+      ...detectMonthName(text, anchorMs),
+      ...detectOrdinalOnly(text, anchorMs),
+      ...detectRelative(text, anchorMs),
+    ];
+
+    for (const m of matches) {
+      const title = clauseAround(text, m.index, m.matched.length);
+      const hasDeadlineWord = DEADLINE_WORDS.some((w) =>
+        title.toLowerCase().includes(w),
+      );
+      let confidence = m.baseConfidence;
+      if (hasDeadlineWord) confidence = Math.min(1, confidence + 0.1);
+      else confidence = Math.max(0, confidence - 0.2);
+
+      const key = `${m.dateMs}|${title.toLowerCase()}`;
+      if (seen.has(key)) continue;
+      seen.add(key);
+
+      out.push({
+        title,
+        dateMs: m.dateMs,
+        allDay: !m.hasTime,
+        sourceSegmentId: seg.id,
+        start: seg.start,
+        raw: m.matched.trim(),
+        confidence: Number(confidence.toFixed(2)),
+      });
+    }
+  }
+
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// Absolute detectors
+// ---------------------------------------------------------------------------
+
+/** ISO `YYYY-MM-DD`, optionally followed by a clock time. */
+function detectIso(text: string, anchorMs: number): RawMatch[] {
+  const out: RawMatch[] = [];
+  const re = /\b(\d{4})-(\d{2})-(\d{2})\b/g;
+  let m: RegExpExecArray | null;
+  while ((m = re.exec(text))) {
+    const year = Number(m[1]);
+    const month = Number(m[2]) - 1;
+    const day = Number(m[3]);
+    if (month < 0 || month > 11 || day < 1 || day > 31) continue;
+    const time = findTimeNear(text, m.index + m[0].length);
+    const dateMs = buildUtc(year, month, day, time);
+    out.push({
+      dateMs,
+      hasTime: !!time,
+      matched: m[0],
+      index: m.index,
+      baseConfidence: 0.9,
+    });
+  }
+  return out;
+}
+
+/**
+ * "may" is also the (very common) modal verb, so a bare "may <number>" — e.g.
+ * "you may 5 minutes", "you may submit 3 drafts" — must NOT be read as the month
+ * May. Only accept it as a month with stronger context: an ordinal on the day
+ * ("May 5th"), an explicit year ("May 5, 2026"), or a date preposition right
+ * before it ("on/by/due/until/before/after May 5").
+ */
+function mayLooksLikeMonth(m: RegExpExecArray, text: string, hasYear: boolean): boolean {
+  if (/\d\s*(?:st|nd|rd|th)\b/i.test(m[0])) return true;
+  if (hasYear) return true;
+  const before = text.slice(Math.max(0, m.index - 10), m.index).toLowerCase();
+  return /\b(?:on|by|due|until|before|after|through)\s*$/.test(before);
+}
+
+/** `March 3`, `March 3, 2026`, `3 March`, `on March 3rd`. */
+function detectMonthName(text: string, anchorMs: number): RawMatch[] {
+  const out: RawMatch[] = [];
+  const monthAlt = Object.keys(MONTHS).join('|');
+
+  // Month-first: "March 3", "March 3rd, 2026"
+  const reMD = new RegExp(
+    `\\b(${monthAlt})\\.?\\s+(\\d{1,2})(?:st|nd|rd|th)?(?:,?\\s+(\\d{4}))?`,
+    'gi',
+  );
+  let m: RegExpExecArray | null;
+  while ((m = reMD.exec(text))) {
+    const month = MONTHS[m[1]!.toLowerCase()];
+    const day = Number(m[2]);
+    if (month === undefined || day < 1 || day > 31) continue;
+    const explicitYear = m[3] ? Number(m[3]) : undefined;
+    if (m[1]!.toLowerCase() === 'may' && !mayLooksLikeMonth(m, text, explicitYear !== undefined)) {
+      continue;
+    }
+    pushResolved(out, text, m, month, day, explicitYear, anchorMs);
+  }
+
+  // Day-first: "3 March", "3rd March 2026"
+  const reDM = new RegExp(
+    `\\b(\\d{1,2})(?:st|nd|rd|th)?\\s+(${monthAlt})\\.?(?:,?\\s+(\\d{4}))?`,
+    'gi',
+  );
+  while ((m = reDM.exec(text))) {
+    const day = Number(m[1]);
+    const month = MONTHS[m[2]!.toLowerCase()];
+    if (month === undefined || day < 1 || day > 31) continue;
+    const explicitYear = m[3] ? Number(m[3]) : undefined;
+    if (m[2]!.toLowerCase() === 'may' && !mayLooksLikeMonth(m, text, explicitYear !== undefined)) {
+      continue;
+    }
+    pushResolved(out, text, m, month, day, explicitYear, anchorMs);
+  }
+
+  return out;
+}
+
+function pushResolved(
+  out: RawMatch[],
+  text: string,
+  m: RegExpExecArray,
+  month: number,
+  day: number,
+  explicitYear: number | undefined,
+  anchorMs: number,
+): void {
+  const time = findTimeNear(text, m.index + m[0].length);
+  let year = explicitYear ?? utcYear(anchorMs);
+  let dateMs = buildUtc(year, month, day, time);
+  if (explicitYear === undefined && dateMs < anchorMs - SIX_MONTHS_MS) {
+    year += 1;
+    dateMs = buildUtc(year, month, day, time);
+  }
+  out.push({
+    dateMs,
+    hasTime: !!time,
+    matched: m[0],
+    index: m.index,
+    baseConfidence: 0.9,
+  });
+}
+
+/** `on the 14th` — day-of-month only, month/year from the anchor. */
+function detectOrdinalOnly(text: string, anchorMs: number): RawMatch[] {
+  const out: RawMatch[] = [];
+  const re = /\bon the (\d{1,2})(?:st|nd|rd|th)\b/gi;
+  let m: RegExpExecArray | null;
+  while ((m = re.exec(text))) {
+    const day = Number(m[1]);
+    if (day < 1 || day > 31) continue;
+    const month = utcMonth(anchorMs);
+    const year = utcYear(anchorMs);
+    const time = findTimeNear(text, m.index + m[0].length);
+    let dateMs = buildUtc(year, month, day, time);
+    // "on the 14th" means the upcoming 14th; if already well past, roll a month.
+    if (dateMs < anchorMs - SIX_MONTHS_MS) {
+      dateMs = buildUtc(year + 1, month, day, time);
+    }
+    out.push({
+      dateMs,
+      hasTime: !!time,
+      matched: m[0],
+      index: m.index,
+      baseConfidence: 0.85,
+    });
+  }
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// Relative detectors (anchored purely to anchorMs)
+// ---------------------------------------------------------------------------
+
+function detectRelative(text: string, anchorMs: number): RawMatch[] {
+  const out: RawMatch[] = [];
+  const dayMs = 1000 * 60 * 60 * 24;
+
+  const add = (offsetDays: number, m: RegExpExecArray) => {
+    const time = findTimeNear(text, m.index + m[0].length);
+    const dateMs = applyTime(anchorMs + offsetDays * dayMs, time);
+    out.push({
+      dateMs,
+      hasTime: !!time,
+      matched: m[0],
+      index: m.index,
+      baseConfidence: 0.6,
+    });
+  };
+
+  let m: RegExpExecArray | null;
+
+  const reToday = /\btoday\b/gi;
+  while ((m = reToday.exec(text))) add(0, m);
+
+  const reTomorrow = /\btomorrow\b/gi;
+  while ((m = reTomorrow.exec(text))) add(1, m);
+
+  const reNextWeek = /\bnext week\b/gi;
+  while ((m = reNextWeek.exec(text))) add(7, m);
+
+  // "in N days" / "in N weeks"
+  const reInN = /\bin (\d{1,3}) (day|days|week|weeks)\b/gi;
+  while ((m = reInN.exec(text))) {
+    const n = Number(m[1]);
+    const unit = m[2]!.toLowerCase();
+    const days = unit.startsWith('week') ? n * 7 : n;
+    add(days, m);
+  }
+
+  // "next Monday".."next Friday" (and weekend days)
+  const reNextDow = /\bnext (sunday|monday|tuesday|wednesday|thursday|friday|saturday)\b/gi;
+  while ((m = reNextDow.exec(text))) {
+    const target = WEEKDAYS[m[1]!.toLowerCase()]!;
+    const anchorDow = utcDow(anchorMs);
+    // Days until the target weekday in the FOLLOWING week (1..7, never 0).
+    let delta = (target - anchorDow + 7) % 7;
+    if (delta === 0) delta = 7;
+    delta += 7; // "next" => skip the current week's occurrence.
+    add(delta, m);
+  }
+
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// Time-of-day extraction
+// ---------------------------------------------------------------------------
+
+/** Look for a clock time shortly after the date phrase (within ~16 chars). */
+function findTimeNear(text: string, from: number): TimeOfDay | undefined {
+  const window = text.slice(from, from + 18);
+  return parseTime(window);
+}
+
+function parseTime(s: string): TimeOfDay | undefined {
+  // "at 3pm", "at 3:30 pm", "3pm"
+  const re12 = /\b(?:at\s+)?(\d{1,2})(?::(\d{2}))?\s*(am|pm)\b/i;
+  const m12 = re12.exec(s);
+  if (m12) {
+    let hour = Number(m12[1]);
+    const minute = m12[2] ? Number(m12[2]) : 0;
+    const mer = m12[3]!.toLowerCase();
+    if (hour === 12) hour = 0;
+    if (mer === 'pm') hour += 12;
+    if (hour >= 0 && hour <= 23 && minute >= 0 && minute <= 59) {
+      return { hour, minute };
+    }
+  }
+  // 24h "14:00", "at 09:30"
+  const re24 = /\b(?:at\s+)?([01]?\d|2[0-3]):([0-5]\d)\b/;
+  const m24 = re24.exec(s);
+  if (m24) {
+    return { hour: Number(m24[1]), minute: Number(m24[2]) };
+  }
+  return undefined;
+}
+
+// ---------------------------------------------------------------------------
+// Pure UTC helpers (constructed only from explicit components / anchorMs)
+// ---------------------------------------------------------------------------
+
+function buildUtc(
+  year: number,
+  month: number,
+  day: number,
+  time: TimeOfDay | undefined,
+): number {
+  return Date.UTC(year, month, day, time?.hour ?? 0, time?.minute ?? 0, 0, 0);
+}
+
+/** Replace the time-of-day on a relative instant, keeping its UTC calendar day. */
+function applyTime(ms: number, time: TimeOfDay | undefined): number {
+  if (!time) return ms;
+  const d = new Date(ms);
+  return Date.UTC(
+    d.getUTCFullYear(),
+    d.getUTCMonth(),
+    d.getUTCDate(),
+    time.hour,
+    time.minute,
+    0,
+    0,
+  );
+}
+
+function utcYear(ms: number): number {
+  return new Date(ms).getUTCFullYear();
+}
+function utcMonth(ms: number): number {
+  return new Date(ms).getUTCMonth();
+}
+function utcDow(ms: number): number {
+  return new Date(ms).getUTCDay();
+}
+
+// ---------------------------------------------------------------------------
+// Clause extraction for titles
+// ---------------------------------------------------------------------------
+
+/** Extract a short clause around the matched date as the event title. */
+function clauseAround(text: string, index: number, length: number): string {
+  // Bound to the surrounding sentence by splitting on sentence punctuation.
+  const before = text.slice(0, index);
+  const after = text.slice(index + length);
+
+  const startBreak = Math.max(
+    before.lastIndexOf('. '),
+    before.lastIndexOf('? '),
+    before.lastIndexOf('! '),
+    before.lastIndexOf('; '),
+    before.lastIndexOf(', '),
+  );
+  const afterBreakRel = firstBreak(after);
+
+  const start = startBreak >= 0 ? startBreak + 2 : 0;
+  const end =
+    index + length + (afterBreakRel >= 0 ? afterBreakRel : after.length);
+
+  return text.slice(start, end).trim().replace(/\s+/g, ' ');
+}
+
+function firstBreak(s: string): number {
+  const idxs = ['. ', '? ', '! ', '; ', ', ']
+    .map((p) => s.indexOf(p))
+    .filter((i) => i >= 0);
+  if (!idxs.length) return -1;
+  return Math.min(...idxs);
+}

src/lib/enrich/ics.test.ts

@@ -0,0 +1,100 @@
+import { describe, it, expect } from 'vitest';
+import { toIcs } from './ics';
+
+const utc = (y: number, mo: number, d: number, h = 0, mi = 0) =>
+  Date.UTC(y, mo, d, h, mi, 0, 0);
+
+describe('toIcs', () => {
+  it('produces a valid empty calendar for []', () => {
+    const ics = toIcs([]);
+    expect(ics).toContain('BEGIN:VCALENDAR');
+    expect(ics).toContain('VERSION:2.0');
+    expect(ics).toContain('END:VCALENDAR');
+    expect(ics).not.toContain('BEGIN:VEVENT');
+  });
+
+  it('uses CRLF line endings', () => {
+    const ics = toIcs([]);
+    expect(ics).toContain('\r\n');
+    expect(ics.endsWith('\r\n')).toBe(true);
+    // No bare LF without a preceding CR.
+    expect(/[^\r]\n/.test(ics)).toBe(false);
+  });
+
+  it('emits an all-day VEVENT with VALUE=DATE', () => {
+    const ics = toIcs([
+      { title: 'Final exam', dateMs: utc(2026, 5, 20), allDay: true },
+    ]);
+    expect(ics).toContain('BEGIN:VEVENT');
+    expect(ics).toContain('DTSTART;VALUE=DATE:20260620');
+    expect(ics).toContain('SUMMARY:Final exam');
+  });
+
+  it('emits a timed VEVENT with a UTC DTSTART', () => {
+    const ics = toIcs([
+      { title: 'Meeting', dateMs: utc(2026, 5, 20, 14, 30), allDay: false },
+    ]);
+    expect(ics).toContain('DTSTART:20260620T143000Z');
+  });
+
+  it('derives DTSTAMP from the event dateMs', () => {
+    const ics = toIcs([
+      { title: 'X', dateMs: utc(2026, 5, 20, 9, 0), allDay: true },
+    ]);
+    expect(ics).toContain('DTSTAMP:20260620T090000Z');
+  });
+
+  it('UID is deterministic (no randomness) and stable across calls', () => {
+    const ev = { title: 'Exam', dateMs: utc(2026, 5, 20), allDay: true };
+    const a = toIcs([ev]);
+    const b = toIcs([ev]);
+    expect(a).toBe(b);
+    const uid = a.split('\r\n').find((l) => l.startsWith('UID:'));
+    expect(uid).toBeDefined();
+    expect(uid).toContain('@wisp');
+  });
+
+  it('changes UID when title or date changes', () => {
+    const a = toIcs([{ title: 'A', dateMs: utc(2026, 5, 20), allDay: true }]);
+    const b = toIcs([{ title: 'B', dateMs: utc(2026, 5, 20), allDay: true }]);
+    const uidA = a.split('\r\n').find((l) => l.startsWith('UID:'));
+    const uidB = b.split('\r\n').find((l) => l.startsWith('UID:'));
+    expect(uidA).not.toBe(uidB);
+  });
+
+  it('escapes comma, semicolon, backslash and newline in SUMMARY', () => {
+    const ics = toIcs([
+      {
+        title: 'a, b; c \\ d\ne',
+        dateMs: utc(2026, 5, 20),
+        allDay: true,
+      },
+    ]);
+    const line = ics
+      .replace(/\r\n /g, '') // unfold first
+      .split('\r\n')
+      .find((l) => l.startsWith('SUMMARY:'));
+    expect(line).toBe('SUMMARY:a\\, b\\; c \\\\ d\\ne');
+  });
+
+  it('folds lines longer than 75 octets with a leading space', () => {
+    const long = 'X'.repeat(200);
+    const ics = toIcs([{ title: long, dateMs: utc(2026, 5, 20), allDay: true }]);
+    for (const line of ics.split('\r\n')) {
+      // Each physical line must be <=75 octets.
+      expect(Buffer.byteLength(line, 'utf8')).toBeLessThanOrEqual(75);
+    }
+    // Unfolding restores the SUMMARY content.
+    const unfolded = ics.replace(/\r\n /g, '');
+    expect(unfolded).toContain(`SUMMARY:${long}`);
+  });
+
+  it('emits one VEVENT per event', () => {
+    const ics = toIcs([
+      { title: 'A', dateMs: utc(2026, 5, 20), allDay: true },
+      { title: 'B', dateMs: utc(2026, 5, 21), allDay: true },
+    ]);
+    const count = ics.split('BEGIN:VEVENT').length - 1;
+    expect(count).toBe(2);
+  });
+});

src/lib/enrich/ics.ts

@@ -0,0 +1,142 @@
+// Pure RFC 5545 iCalendar serialization. Deterministic (no randomness, no
+// wall-clock): UID and DTSTAMP are derived from the event itself, so the same
+// input always yields byte-identical output. CRLF line endings, 75-octet
+// folding.
+
+interface IcsEvent {
+  title: string;
+  dateMs: number;
+  allDay: boolean;
+}
+
+/** Serialize events to a VCALENDAR string. Valid (and non-empty) even for []. */
+export function toIcs(events: IcsEvent[]): string {
+  const lines: string[] = [
+    'BEGIN:VCALENDAR',
+    'VERSION:2.0',
+    'PRODID:-//Wisp//Enrich//EN',
+    'CALSCALE:GREGORIAN',
+  ];
+
+  for (const ev of events) {
+    const stamp = formatUtc(ev.dateMs);
+    lines.push('BEGIN:VEVENT');
+    lines.push(`UID:${uidFor(ev)}`);
+    lines.push(`DTSTAMP:${stamp}`);
+    if (ev.allDay) {
+      lines.push(`DTSTART;VALUE=DATE:${formatDate(ev.dateMs)}`);
+    } else {
+      lines.push(`DTSTART:${stamp}`);
+    }
+    lines.push(`SUMMARY:${escapeText(ev.title)}`);
+    lines.push('END:VEVENT');
+  }
+
+  lines.push('END:VCALENDAR');
+
+  return lines.map(foldLine).join('\r\n') + '\r\n';
+}
+
+// ---------------------------------------------------------------------------
+// Deterministic UID derivation (FNV-1a hash of title + dateMs)
+// ---------------------------------------------------------------------------
+
+function uidFor(ev: IcsEvent): string {
+  const h = fnv1a(`${ev.title}|${ev.dateMs}`);
+  return `${h}@wisp`;
+}
+
+function fnv1a(s: string): string {
+  let hash = 0x811c9dc5;
+  for (let i = 0; i < s.length; i++) {
+    hash ^= s.charCodeAt(i);
+    // 32-bit FNV prime multiply via shifts, kept unsigned.
+    hash = Math.imul(hash, 0x01000193) >>> 0;
+  }
+  return hash.toString(16).padStart(8, '0');
+}
+
+// ---------------------------------------------------------------------------
+// Date formatting from dateMs only (UTC)
+// ---------------------------------------------------------------------------
+
+function formatDate(ms: number): string {
+  const d = new Date(ms);
+  return (
+    pad(d.getUTCFullYear(), 4) + pad(d.getUTCMonth() + 1) + pad(d.getUTCDate())
+  );
+}
+
+function formatUtc(ms: number): string {
+  const d = new Date(ms);
+  return (
+    pad(d.getUTCFullYear(), 4) +
+    pad(d.getUTCMonth() + 1) +
+    pad(d.getUTCDate()) +
+    'T' +
+    pad(d.getUTCHours()) +
+    pad(d.getUTCMinutes()) +
+    pad(d.getUTCSeconds()) +
+    'Z'
+  );
+}
+
+function pad(n: number, width = 2): string {
+  return String(n).padStart(width, '0');
+}
+
+// ---------------------------------------------------------------------------
+// RFC 5545 text escaping + line folding
+// ---------------------------------------------------------------------------
+
+/** Escape backslash, semicolon, comma, and newlines per RFC 5545 §3.3.11. */
+function escapeText(s: string): string {
+  return s
+    .replace(/\\/g, '\\\\')
+    .replace(/;/g, '\\;')
+    .replace(/,/g, '\\,')
+    .replace(/\r\n|\n|\r/g, '\\n');
+}
+
+/**
+ * Fold a content line to <=75 octets, continuation lines starting with a
+ * single space. UTF-8 aware: never splits a multi-byte character across the
+ * fold boundary, so every emitted line is valid UTF-8.
+ */
+function foldLine(line: string): string {
+  if (utf8Len(line) <= 75) return line;
+
+  const chunks: string[] = [];
+  // First line: 75 octets. Continuation lines: 74 octets (1 leading space).
+  let limit = 75;
+  let current = '';
+  let currentBytes = 0;
+
+  for (const ch of line) {
+    const w = utf8Len(ch);
+    if (currentBytes + w > limit) {
+      chunks.push(current);
+      current = '';
+      currentBytes = 0;
+      limit = 74;
+    }
+    current += ch;
+    currentBytes += w;
+  }
+  if (current) chunks.push(current);
+
+  return chunks.map((c, i) => (i === 0 ? c : ' ' + c)).join('\r\n');
+}
+
+/** Number of UTF-8 octets a string encodes to. */
+function utf8Len(s: string): number {
+  let n = 0;
+  for (const ch of s) {
+    const cp = ch.codePointAt(0)!;
+    if (cp < 0x80) n += 1;
+    else if (cp < 0x800) n += 2;
+    else if (cp < 0x10000) n += 3;
+    else n += 4;
+  }
+  return n;
+}

src/lib/enrich/index.ts

@@ -0,0 +1,11 @@
+// Phase 4 enrichment barrel: pure detection/serialization + network lookups.
+// Pure modules contain NO transcript text in any outbound URL; the network
+// module sends only the bare identifier/topic. See each file's header.
+
+export * from './types';
+export * from './dates';
+export * from './citations';
+export * from './ics';
+export * from './bib';
+export * from './links';
+export * from './lookup';

src/lib/enrich/links.test.ts

@@ -0,0 +1,77 @@
+import { describe, it, expect } from 'vitest';
+import { lookupLinksFor, bookLinksFor } from './links';
+import type { Citation, RefMetadata } from './types';
+
+const cite = (kind: Citation['kind'], value: string): Citation => ({
+  kind,
+  value,
+  raw: value,
+});
+
+describe('lookupLinksFor', () => {
+  it('builds a DOI resolver link', () => {
+    const l = lookupLinksFor(cite('doi', '10.1038/nphys1170'));
+    expect(l.doi).toBe('https://doi.org/10.1038%2Fnphys1170');
+  });
+
+  it('builds an arXiv abstract link', () => {
+    const l = lookupLinksFor(cite('arxiv', '1706.03762'));
+    expect(l.arxiv).toBe('https://arxiv.org/abs/1706.03762');
+  });
+
+  it('points openLibrary at the ISBN page for isbn citations', () => {
+    const l = lookupLinksFor(cite('isbn', '9780134685991'));
+    expect(l.openLibrary).toBe('https://openlibrary.org/isbn/9780134685991');
+  });
+
+  it('always includes wikipedia/googleBooks search links', () => {
+    const l = lookupLinksFor(cite('author-year', 'Vaswani (2017)'));
+    expect(l.wikipedia).toContain(
+      'https://en.wikipedia.org/wiki/Special:Search?search=',
+    );
+    expect(l.googleBooks).toContain('https://www.google.com/search?tbm=bks&q=');
+  });
+
+  it('always includes annasArchive and libgen search links', () => {
+    const l = lookupLinksFor(cite('doi', '10.1/x'));
+    expect(l.annasArchive).toContain('https://annas-archive.org/search?q=');
+    expect(l.libgen).toContain('https://libgen.is/search.php?req=');
+  });
+
+  it('prefers resolved metadata title for the search query, URL-encoded', () => {
+    const meta: RefMetadata = {
+      title: 'Attention Is All You Need',
+      source: 'Crossref',
+    };
+    const l = lookupLinksFor(cite('arxiv', '1706.03762'), meta);
+    const q = encodeURIComponent('Attention Is All You Need');
+    expect(l.wikipedia).toContain(q);
+    expect(l.annasArchive).toContain(q);
+  });
+
+  it('falls back to the citation value when no metadata title', () => {
+    const l = lookupLinksFor(cite('author-year', 'Smith et al. (2021)'));
+    const q = encodeURIComponent('Smith et al. (2021)');
+    expect(l.googleBooks).toContain(q);
+  });
+});
+
+describe('bookLinksFor', () => {
+  it('builds the full search set from a free-text query (no doi/arxiv)', () => {
+    const l = bookLinksFor('Introduction to Algorithms');
+    const q = encodeURIComponent('Introduction to Algorithms');
+    expect(l.doi).toBeUndefined();
+    expect(l.arxiv).toBeUndefined();
+    expect(l.openLibrary).toBe(`https://openlibrary.org/search?q=${q}`);
+    expect(l.wikipedia).toContain(q);
+    expect(l.googleBooks).toContain(q);
+    expect(l.annasArchive).toContain(q);
+    expect(l.libgen).toContain(q);
+  });
+
+  it('URL-encodes special characters', () => {
+    const l = bookLinksFor('C&A: a/b test');
+    expect(l.libgen).toContain(encodeURIComponent('C&A: a/b test'));
+    expect(l.libgen).not.toContain(' ');
+  });
+});

src/lib/enrich/links.ts

@@ -0,0 +1,47 @@
+// Pure outbound-link builders. NO network here — these only construct URLs.
+// Privacy: the only data placed in a URL is the extracted citation value or a
+// resolved title — never transcript text. Anna's Archive / LibGen appear as
+// SEARCH urls only; nothing in this app fetches or proxies them.
+
+import type { Citation, LookupLinks, RefMetadata } from './types';
+
+/** Build lookup links for a detected citation, optionally enriched by metadata. */
+export function lookupLinksFor(c: Citation, meta?: RefMetadata): LookupLinks {
+  const query = (meta?.title || c.value).trim();
+
+  const links: LookupLinks = {
+    ...searchLinks(query),
+  };
+
+  if (c.kind === 'doi') {
+    links.doi = `https://doi.org/${encodeURIComponent(c.value)}`;
+  }
+  if (c.kind === 'arxiv') {
+    links.arxiv = `https://arxiv.org/abs/${encodeURIComponent(c.value)}`;
+  }
+  if (c.kind === 'isbn') {
+    // Direct ISBN page beats a free-text Open Library search.
+    links.openLibrary = `https://openlibrary.org/isbn/${encodeURIComponent(
+      c.value,
+    )}`;
+  }
+
+  return links;
+}
+
+/** Build the same set of links from a free-text query (no DOI/arXiv). */
+export function bookLinksFor(query: string): LookupLinks {
+  return searchLinks(query.trim());
+}
+
+/** The query-driven search links shared by both entry points. */
+function searchLinks(query: string): LookupLinks {
+  const q = encodeURIComponent(query);
+  return {
+    wikipedia: `https://en.wikipedia.org/wiki/Special:Search?search=${q}`,
+    googleBooks: `https://www.google.com/search?tbm=bks&q=${q}`,
+    openLibrary: `https://openlibrary.org/search?q=${q}`,
+    annasArchive: `https://annas-archive.org/search?q=${q}`,
+    libgen: `https://libgen.is/search.php?req=${q}`,
+  };
+}

src/lib/enrich/lookup.ts

@@ -0,0 +1,290 @@
+// NETWORK metadata client for Phase 4 enrichment.
+//
+// PRIVACY: every request below sends ONLY the extracted citation id (DOI /
+// arXiv id / ISBN) or the topic string — never any transcript text. Detection
+// already happened locally; this module just resolves a bare identifier to
+// public metadata.
+//
+// All endpoints are CORS-friendly, legitimate, metadata-only APIs (Crossref,
+// OpenAlex, Open Library, Wikipedia REST). Their responses include the
+// `Access-Control-Allow-Origin` header, so they are readable even under our
+// COEP `require-corp` / COOP `same-origin` isolation (CORS responses satisfy
+// require-corp without needing CORP headers). We never fetch or proxy any
+// shadow-library content here — those are search-URL only and live in links.ts.
+//
+// Every function is defensive: ~8s AbortController timeout, try/catch that
+// resolves to `undefined`, and never throws. Results are memoised in an
+// in-memory Map for the session.
+
+import type { Citation, RefMetadata } from './types';
+
+/** Per-session metadata cache, keyed by `${kind}:${value}` / `wiki:${topic}`. */
+const cache = new Map<string, RefMetadata | undefined>();
+const wikiCache = new Map<
+  string,
+  { title: string; extract: string; url: string } | undefined
+>();
+
+const TIMEOUT_MS = 8000;
+
+/** GET JSON with an abort timeout. Returns parsed JSON or undefined on any
+ *  failure (network, non-2xx, timeout, parse). Never throws. */
+async function fetchJson(url: string): Promise<unknown | undefined> {
+  const controller = new AbortController();
+  const timer = setTimeout(() => controller.abort(), TIMEOUT_MS);
+  try {
+    const res = await fetch(url, {
+      method: 'GET',
+      headers: { Accept: 'application/json' },
+      signal: controller.signal,
+    });
+    if (!res.ok) return undefined;
+    return (await res.json()) as unknown;
+  } catch {
+    return undefined;
+  } finally {
+    clearTimeout(timer);
+  }
+}
+
+/** Narrow an unknown to a plain record for safe property access. */
+function asRecord(v: unknown): Record<string, unknown> | undefined {
+  return typeof v === 'object' && v !== null
+    ? (v as Record<string, unknown>)
+    : undefined;
+}
+
+function asString(v: unknown): string | undefined {
+  return typeof v === 'string' && v.length > 0 ? v : undefined;
+}
+
+function asNumber(v: unknown): number | undefined {
+  return typeof v === 'number' && Number.isFinite(v) ? v : undefined;
+}
+
+// ---------------------------------------------------------------------------
+// Crossref (DOI)
+// ---------------------------------------------------------------------------
+
+function parseCrossref(json: unknown, value: string): RefMetadata | undefined {
+  const root = asRecord(json);
+  const message = asRecord(root?.message);
+  if (!message) return undefined;
+
+  const titleArr = Array.isArray(message.title) ? message.title : [];
+  const title = asString(titleArr[0]);
+
+  const authors: string[] = [];
+  if (Array.isArray(message.author)) {
+    for (const a of message.author) {
+      const rec = asRecord(a);
+      if (!rec) continue;
+      const given = asString(rec.given);
+      const family = asString(rec.family);
+      const name = [given, family].filter(Boolean).join(' ').trim();
+      if (name) authors.push(name);
+    }
+  }
+
+  // Year: prefer published, then published-print/online, then issued.
+  const year =
+    crossrefYear(message.published) ??
+    crossrefYear(message['published-print']) ??
+    crossrefYear(message['published-online']) ??
+    crossrefYear(message.issued);
+
+  return {
+    title,
+    authors: authors.length ? authors : undefined,
+    year,
+    source: 'Crossref',
+    url: `https://doi.org/${value}`,
+  };
+}
+
+/** Crossref date parts live at `<field>['date-parts'][0][0]`. */
+function crossrefYear(field: unknown): number | undefined {
+  const rec = asRecord(field);
+  const parts = rec?.['date-parts'];
+  if (!Array.isArray(parts)) return undefined;
+  const first = parts[0];
+  if (!Array.isArray(first)) return undefined;
+  return asNumber(first[0]);
+}
+
+// ---------------------------------------------------------------------------
+// OpenAlex (arXiv, and DOI fallback)
+// ---------------------------------------------------------------------------
+
+function parseOpenAlex(
+  json: unknown,
+  source: string,
+  url: string,
+): RefMetadata | undefined {
+  const work = asRecord(json);
+  if (!work) return undefined;
+
+  const title = asString(work.title) ?? asString(work.display_name);
+
+  const authors: string[] = [];
+  if (Array.isArray(work.authorships)) {
+    for (const a of work.authorships) {
+      const rec = asRecord(a);
+      const author = asRecord(rec?.author);
+      const name = asString(author?.display_name);
+      if (name) authors.push(name);
+    }
+  }
+
+  const year = asNumber(work.publication_year);
+
+  // OpenAlex returns an error object (with no title) for unknown ids.
+  if (!title && !authors.length && year === undefined) return undefined;
+
+  return {
+    title,
+    authors: authors.length ? authors : undefined,
+    year,
+    source,
+    url,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Open Library (ISBN)
+// ---------------------------------------------------------------------------
+
+function parseOpenLibrary(
+  json: unknown,
+  value: string,
+): RefMetadata | undefined {
+  const book = asRecord(json);
+  if (!book) return undefined;
+
+  const title = asString(book.title);
+
+  // Open Library's /isbn/{}.json gives `by_statement` (a free-text credit)
+  // rather than structured authors; surface it as a single author when present.
+  const byStatement = asString(book.by_statement);
+  const authors = byStatement ? [byStatement] : undefined;
+
+  // publish_date is free-text (e.g. "March 2004"); pull a 4-digit year out.
+  let year: number | undefined;
+  const publishDate = asString(book.publish_date);
+  if (publishDate) {
+    const m = publishDate.match(/\b(\d{4})\b/);
+    if (m) year = asNumber(Number(m[1]));
+  }
+
+  if (!title && !authors && year === undefined) return undefined;
+
+  return {
+    title,
+    authors,
+    year,
+    source: 'Open Library',
+    url: `https://openlibrary.org/isbn/${value}`,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Public API
+// ---------------------------------------------------------------------------
+
+/**
+ * Resolve a detected citation to public metadata.
+ *
+ * Sends ONLY the citation's identifier (DOI / arXiv id / ISBN) to a
+ * metadata-only API — never any transcript text. Returns `undefined` on any
+ * failure or for `author-year` citations (the UI offers search links instead).
+ * Never throws.
+ */
+export async function resolveCitation(
+  c: Citation,
+): Promise<RefMetadata | undefined> {
+  const key = `${c.kind}:${c.value}`;
+  if (cache.has(key)) return cache.get(key);
+
+  let result: RefMetadata | undefined;
+
+  switch (c.kind) {
+    case 'doi': {
+      // Crossref is the authoritative DOI registry and is CORS-enabled.
+      const json = await fetchJson(
+        `https://api.crossref.org/works/${encodeURIComponent(c.value)}`,
+      );
+      result = parseCrossref(json, c.value);
+      // Fallback to OpenAlex if Crossref had nothing (e.g. DataCite DOI).
+      if (!result) {
+        const alt = await fetchJson(
+          `https://api.openalex.org/works/doi:${encodeURIComponent(c.value)}`,
+        );
+        result = parseOpenAlex(alt, 'OpenAlex', `https://doi.org/${c.value}`);
+      }
+      break;
+    }
+    case 'arxiv': {
+      // OpenAlex resolves arXiv ids over CORS, avoiding arXiv's
+      // non-CORS Atom export API.
+      const json = await fetchJson(
+        `https://api.openalex.org/works/arxiv:${encodeURIComponent(c.value)}`,
+      );
+      result = parseOpenAlex(
+        json,
+        'OpenAlex',
+        `https://arxiv.org/abs/${c.value}`,
+      );
+      break;
+    }
+    case 'isbn': {
+      const json = await fetchJson(
+        `https://openlibrary.org/isbn/${encodeURIComponent(c.value)}.json`,
+      );
+      result = parseOpenLibrary(json, c.value);
+      break;
+    }
+    case 'author-year':
+      // No reliable id to resolve; UI falls back to search links.
+      result = undefined;
+      break;
+  }
+
+  cache.set(key, result);
+  return result;
+}
+
+/**
+ * Fetch a one-paragraph Wikipedia summary for a topic.
+ *
+ * Sends ONLY the topic string to Wikipedia's REST summary endpoint (CORS-
+ * enabled) — never any transcript text. Returns `undefined` on 404 / any
+ * failure. Never throws.
+ */
+export async function wikipediaSummary(
+  topic: string,
+): Promise<{ title: string; extract: string; url: string } | undefined> {
+  const key = `wiki:${topic}`;
+  if (wikiCache.has(key)) return wikiCache.get(key);
+
+  let result: { title: string; extract: string; url: string } | undefined;
+
+  const json = await fetchJson(
+    `https://en.wikipedia.org/api/rest_v1/page/summary/${encodeURIComponent(
+      topic,
+    )}`,
+  );
+  const root = asRecord(json);
+  if (root) {
+    const title = asString(root.title);
+    const extract = asString(root.extract);
+    const contentUrls = asRecord(root.content_urls);
+    const desktop = asRecord(contentUrls?.desktop);
+    const url = asString(desktop?.page);
+    if (title && extract && url) {
+      result = { title, extract, url };
+    }
+  }
+
+  wikiCache.set(key, result);
+  return result;
+}

src/lib/enrich/types.ts

@@ -0,0 +1,55 @@
+// Shared types for Phase 4 enrichment: dates -> calendar, references -> lookups.
+// All detection is pure over Segment text; network lookups send only the
+// extracted topic/DOI/ISBN string — never the transcript itself.
+
+/** A detected date/deadline the user can confirm into their calendar. */
+export interface CandidateEvent {
+  /** Short event title (the surrounding clause, trimmed). */
+  title: string;
+  /** Resolved absolute time, ms since epoch. */
+  dateMs: number;
+  /** All-day if no specific time was detected. */
+  allDay: boolean;
+  /** Source segment for jump-to-lecture. */
+  sourceSegmentId?: string;
+  /** Audio start time of the source segment. */
+  start?: number;
+  /** The matched text. */
+  raw: string;
+  /** 0..1 detection confidence (relative dates are lower). */
+  confidence: number;
+}
+
+/** A detected citation/reference in the lecture. */
+export interface Citation {
+  kind: 'doi' | 'arxiv' | 'isbn' | 'author-year';
+  /** Normalized id (DOI, arXiv id, ISBN-13/10) or the raw author-year string. */
+  value: string;
+  sourceSegmentId?: string;
+  start?: number;
+  raw: string;
+}
+
+/** Resolved metadata for a citation, from a legitimate metadata API. */
+export interface RefMetadata {
+  title?: string;
+  authors?: string[];
+  year?: number;
+  /** Which API answered, e.g. 'Crossref' | 'OpenAlex' | 'Open Library'. */
+  source: string;
+  /** Canonical URL (DOI resolver, work page, etc.). */
+  url?: string;
+}
+
+/** Outbound links for finding a reference/book. Legit sources first; the
+ *  shadow-library links (Anna's Archive / LibGen) are search URLs only. */
+export interface LookupLinks {
+  doi?: string;
+  arxiv?: string;
+  openAlex?: string;
+  openLibrary?: string;
+  googleBooks?: string;
+  wikipedia?: string;
+  annasArchive?: string;
+  libgen?: string;
+}

src/lib/errorMessage.ts

@@ -0,0 +1,24 @@
+// Extract a human-readable message from an unknown thrown value.
+//
+// `String(err)` on a non-Error object yields the useless "[object Object]",
+// which is exactly what surfaced when native modules (whisper.rn, the audio
+// decoder, file system) reject with a plain `{ message, code }`-shaped object
+// instead of an `Error`. This pulls the most useful text out of whatever was
+// thrown so the UI shows something actionable.
+export function errorMessage(err: unknown): string {
+  if (err instanceof Error) return err.message;
+  if (typeof err === 'string') return err;
+  if (err && typeof err === 'object') {
+    const o = err as { message?: unknown; reason?: unknown; code?: unknown };
+    const parts = [o.message, o.reason, o.code]
+      .filter((x) => x != null && x !== '')
+      .map(String);
+    if (parts.length > 0) return parts.join(' — ');
+    try {
+      return JSON.stringify(err);
+    } catch {
+      /* fall through to String() for circular/unserializable objects */
+    }
+  }
+  return String(err);
+}

src/lib/generation/cloud.ts

@@ -0,0 +1,95 @@
+// Bring-your-own-key cloud generation engine (OpenAI-compatible chat
+// completions). Used when the user has supplied a CloudConfig (baseUrl, apiKey,
+// model). Works against OpenAI itself or any OpenAI-compatible endpoint.
+//
+// PRIVACY: this engine sends ONLY the system prompt + the user's question +
+// the retrieved lecture snippets (the caller assembles these via prompt.ts).
+// Raw audio and full transcripts are NEVER sent — that is a hard Phase 5 rule.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+import type { CloudConfig, GenerationEngine, GenOptions } from './engine';
+
+interface ChatMessage {
+  role: 'system' | 'user';
+  content: string;
+}
+
+interface ChatCompletionResponse {
+  choices?: { message?: { content?: string } }[];
+}
+
+export function createCloudEngine(cfg: CloudConfig): GenerationEngine {
+  // Normalize: drop a trailing slash so we can append the path cleanly.
+  const base = cfg.baseUrl.replace(/\/+$/, '');
+
+  return {
+    kind: 'cloud',
+    label: 'Cloud (your key)',
+
+    async isAvailable(): Promise<boolean> {
+      return !!cfg.apiKey;
+    },
+
+    // Cloud models need no local download; nothing to load.
+    async loadModel(): Promise<void> {
+      /* no-op */
+    },
+
+    isLoaded(): boolean {
+      return true;
+    },
+
+    async generate(prompt: string, opts?: GenOptions): Promise<string> {
+      if (!cfg.apiKey) {
+        throw new Error('Cloud generation requires an API key.');
+      }
+
+      const messages: ChatMessage[] = [];
+      if (opts?.system) messages.push({ role: 'system', content: opts.system });
+      messages.push({ role: 'user', content: prompt });
+
+      let res: Response;
+      try {
+        res = await fetch(`${base}/chat/completions`, {
+          method: 'POST',
+          headers: {
+            'Content-Type': 'application/json',
+            Authorization: `Bearer ${cfg.apiKey}`,
+          },
+          body: JSON.stringify({
+            model: cfg.model,
+            messages,
+            max_tokens: opts?.maxTokens ?? 512,
+          }),
+          signal: opts?.signal,
+        });
+      } catch (err) {
+        if (err instanceof DOMException && err.name === 'AbortError') throw err;
+        throw new Error(
+          `Cloud request failed: ${err instanceof Error ? err.message : String(err)}`,
+        );
+      }
+
+      if (!res.ok) {
+        const detail = await res.text().catch(() => '');
+        throw new Error(
+          `Cloud request failed (${res.status} ${res.statusText})${detail ? `: ${detail}` : ''}`,
+        );
+      }
+
+      let data: ChatCompletionResponse;
+      try {
+        data = (await res.json()) as ChatCompletionResponse;
+      } catch {
+        throw new Error('Cloud response was not valid JSON.');
+      }
+
+      const content = data.choices?.[0]?.message?.content;
+      if (typeof content !== 'string') {
+        throw new Error('Cloud response did not contain a message.');
+      }
+      return content;
+    },
+  };
+}

src/lib/generation/engine.ts

@@ -0,0 +1,62 @@
+// Optional generative layer (ROADMAP Phase 5). STRICTLY opt-in and gated: a
+// local WebLLM model (WebGPU) or a bring-your-own-key cloud model. When neither
+// is available the app falls back to the deterministic Phase 1/3 features — the
+// LLM is never a hard dependency, and raw audio/transcripts are NEVER sent to a
+// cloud: only the user's question + the retrieved snippets go out (BYO-key path).
+
+export interface GenOptions {
+  system?: string;
+  maxTokens?: number;
+  signal?: AbortSignal;
+  /** Streamed token callback (optional). */
+  onToken?: (delta: string) => void;
+}
+
+export interface GenerationEngine {
+  /** Which backend this is. */
+  readonly kind: 'webllm' | 'cloud' | 'none';
+  /** Human label, e.g. 'On-device (Qwen2.5-1.5B)' or 'Cloud (your key)'. */
+  readonly label: string;
+  /** Whether this engine can run here right now (WebGPU present / key set). */
+  isAvailable(): Promise<boolean>;
+  /** Load/prepare the model. onProgress in [0,1] (for the WebLLM download). */
+  loadModel(onProgress?: (p: number) => void): Promise<void>;
+  isLoaded(): boolean;
+  /** Generate a completion for `prompt`. */
+  generate(prompt: string, opts?: GenOptions): Promise<string>;
+}
+
+/** Cloud (BYO-key) configuration — OpenAI-compatible chat completions. */
+export interface CloudConfig {
+  /** e.g. https://api.openai.com/v1 (or any OpenAI-compatible base). */
+  baseUrl: string;
+  apiKey: string;
+  model: string;
+}
+
+/** Default local WebLLM model (small instruct, q4) — pulled from the CDN. */
+export const WEBLLM_MODEL = 'Qwen2.5-1.5B-Instruct-q4f16_1-MLC';
+
+// ---------------------------------------------------------------------------
+// RAG ("ask your lectures") result shape — produced by rag.ts.
+// ---------------------------------------------------------------------------
+
+export interface RagCitation {
+  transcriptId: string;
+  segmentId?: string;
+  start: number;
+  text: string;
+  /** Retrieval score from Phase 1 search. */
+  score: number;
+}
+
+export interface RagAnswer {
+  /** The generated answer (grounded in the citations). */
+  answer: string;
+  /** The lecture moments the answer is based on (always shown verbatim). */
+  citations: RagCitation[];
+  /** False when no relevant lecture content was retrieved (we refuse to answer). */
+  grounded: boolean;
+  /** True when produced by an LLM; false when this is the search-only fallback. */
+  generated: boolean;
+}

src/lib/generation/engineImpl.native.ts

@@ -0,0 +1,32 @@
+// NATIVE on-device generation stub. Running a local LLM on device (e.g. via
+// llama.rn / MLC's native runtime) is a Phase 5 follow-up; until then the native
+// webllm engine reports unavailable so callers transparently fall back to the
+// search-only path (or a BYO-key cloud model). It NEVER produces a fake answer.
+//
+// This module is selected by Metro for native (.native.ts) and is never imported
+// by any vitest test.
+
+import type { GenerationEngine } from './engine';
+
+const NOT_AVAILABLE = 'On-device generation is not available on native yet';
+
+export const webllm: GenerationEngine = {
+  kind: 'webllm',
+  label: 'On-device (Qwen2.5-1.5B)',
+
+  async isAvailable(): Promise<boolean> {
+    return false;
+  },
+
+  async loadModel(): Promise<void> {
+    throw new Error(NOT_AVAILABLE);
+  },
+
+  isLoaded(): boolean {
+    return false;
+  },
+
+  async generate(): Promise<string> {
+    throw new Error(NOT_AVAILABLE);
+  },
+};

src/lib/generation/engineImpl.ts

@@ -0,0 +1,3 @@
+// Base resolver for TypeScript. Metro picks engineImpl.web.ts / engineImpl.native.ts
+// by platform extension at build time; tsc resolves this file (defaults to web).
+export { webllm } from './engineImpl.web';

src/lib/generation/engineImpl.web.ts

@@ -0,0 +1,139 @@
+// WEB-ONLY on-device generation engine, backed by @mlc-ai/web-llm running a
+// small instruct model (Qwen2.5-1.5B, q4) entirely in the browser via WebGPU.
+//
+// WHY WE LOAD IT FROM A CDN AT RUNTIME (not a static import):
+// web-llm ships WASM and uses dynamic imports that Metro (Expo's web bundler)
+// cannot statically bundle. As with transcription/engineImpl.web.ts, we never
+// let Metro see the package: we load the ESM build from a CDN at runtime via a
+// dynamic import hidden behind `new Function` (so Metro's static analyzer can't
+// trip over it). The browser resolves it natively. This keeps the JS bundle
+// small — the model itself is a ~1GB download, cached by the browser after the
+// first run — and means web-llm is NEVER bundled.
+//
+// Requires WebGPU. On a device without a WebGPU adapter, isAvailable() is false
+// and callers fall back to the search-only path (no fake answers).
+//
+// This module is web-only and is NEVER imported by any vitest test.
+
+import { WEBLLM_MODEL } from './engine';
+import type { GenerationEngine, GenOptions } from './engine';
+
+// Pin the web-llm ESM build we load at runtime.
+const WEBLLM_CDN = 'https://esm.run/@mlc-ai/web-llm';
+
+// `new Function` hides the dynamic import() specifier from Metro's bundler so it
+// never tries to resolve/transform web-llm or its WASM.
+const runtimeImport = new Function('u', 'return import(u)') as (u: string) => Promise<WebLlmModule>;
+
+// Minimal structural types for the bits of web-llm we use.
+interface InitProgressReport {
+  progress: number;
+  text?: string;
+}
+interface ChatMessage {
+  role: 'system' | 'user' | 'assistant';
+  content: string;
+}
+interface CompletionChunk {
+  choices: { delta: { content?: string } }[];
+}
+interface CompletionResponse {
+  choices: { message: { content: string } }[];
+}
+interface ChatCompletions {
+  create(req: {
+    messages: ChatMessage[];
+    stream: false;
+    max_tokens: number;
+  }): Promise<CompletionResponse>;
+  create(req: {
+    messages: ChatMessage[];
+    stream: true;
+    max_tokens: number;
+  }): Promise<AsyncIterable<CompletionChunk>>;
+}
+interface MLCEngine {
+  chat: { completions: ChatCompletions };
+}
+interface WebLlmModule {
+  CreateMLCEngine(
+    model: string,
+    opts?: { initProgressCallback?: (p: InitProgressReport) => void },
+  ): Promise<MLCEngine>;
+}
+
+let libPromise: Promise<WebLlmModule> | null = null;
+function lib(): Promise<WebLlmModule> {
+  if (!libPromise) libPromise = runtimeImport(WEBLLM_CDN);
+  return libPromise;
+}
+
+let engineInstance: MLCEngine | null = null;
+
+async function hasWebGpu(): Promise<boolean> {
+  try {
+    if (typeof navigator === 'undefined' || !('gpu' in navigator)) return false;
+    const gpu = (navigator as { gpu?: { requestAdapter(): Promise<unknown> } }).gpu;
+    const adapter = await gpu?.requestAdapter();
+    return adapter != null;
+  } catch {
+    return false;
+  }
+}
+
+export const webllm: GenerationEngine = {
+  kind: 'webllm',
+  label: 'On-device (Qwen2.5-1.5B)',
+
+  async isAvailable(): Promise<boolean> {
+    return hasWebGpu();
+  },
+
+  async loadModel(onProgress?: (p: number) => void): Promise<void> {
+    if (engineInstance) return; // idempotent
+    const { CreateMLCEngine } = await lib();
+    engineInstance = await CreateMLCEngine(WEBLLM_MODEL, {
+      initProgressCallback: (p) => onProgress?.(p.progress),
+    });
+  },
+
+  isLoaded(): boolean {
+    return engineInstance != null;
+  },
+
+  async generate(prompt: string, opts?: GenOptions): Promise<string> {
+    if (!engineInstance) {
+      throw new Error('On-device model is not loaded; call loadModel() first.');
+    }
+    const messages: ChatMessage[] = [];
+    if (opts?.system) messages.push({ role: 'system', content: opts.system });
+    messages.push({ role: 'user', content: prompt });
+
+    const maxTokens = opts?.maxTokens ?? 512;
+
+    if (opts?.onToken) {
+      // Stream: surface each delta and accumulate the full text to return.
+      const stream = await engineInstance.chat.completions.create({
+        messages,
+        stream: true,
+        max_tokens: maxTokens,
+      });
+      let full = '';
+      for await (const chunk of stream) {
+        const delta = chunk.choices[0]?.delta?.content ?? '';
+        if (delta) {
+          full += delta;
+          opts.onToken(delta);
+        }
+      }
+      return full;
+    }
+
+    const res = await engineInstance.chat.completions.create({
+      messages,
+      stream: false,
+      max_tokens: maxTokens,
+    });
+    return res.choices[0]?.message?.content ?? '';
+  },
+};

src/lib/generation/index.ts

@@ -0,0 +1,51 @@
+// Public entry point for the optional generative ("ask your lectures") layer.
+//
+// `./engineImpl` is resolved by Metro to engineImpl.web.ts or
+// engineImpl.native.ts by platform extension; the base engineImpl.ts re-export
+// (web) is what TypeScript resolves for typechecking. Consumers call
+// getGenerationEngine() and stay platform-agnostic.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+import type { CloudConfig, GenerationEngine } from './engine';
+import { createCloudEngine } from './cloud';
+import { webllm } from './engineImpl';
+
+/**
+ * An engine that does nothing useful — represents "no generation backend here".
+ * isAvailable() is always false; generate()/loadModel() throw. Callers should
+ * check isAvailable() and fall back to the search-only path (no fake answers).
+ */
+export const noneEngine: GenerationEngine = {
+  kind: 'none',
+  label: 'No model',
+  async isAvailable(): Promise<boolean> {
+    return false;
+  },
+  async loadModel(): Promise<void> {
+    throw new Error('No generation engine is available.');
+  },
+  isLoaded(): boolean {
+    return false;
+  },
+  async generate(): Promise<string> {
+    throw new Error('No generation engine is available.');
+  },
+};
+
+/**
+ * Pick the generation engine.
+ *
+ * - If a cloud config WITH an apiKey is given, use the BYO-key cloud engine.
+ * - Otherwise return the platform on-device (webllm) engine. On a device without
+ *   WebGPU (web) or on native, that engine's isAvailable() resolves false, which
+ *   is how the "none" state is represented in practice — callers MUST check
+ *   isAvailable() before loading/generating, and fall back to search-only.
+ */
+export function getGenerationEngine(cloud?: CloudConfig): GenerationEngine {
+  if (cloud?.apiKey) return createCloudEngine(cloud);
+  return webllm;
+}
+
+export { createCloudEngine, webllm };
+export * from './engine';

src/lib/generation/prompt.test.ts

@@ -0,0 +1,45 @@
+import { describe, it, expect } from 'vitest';
+import { buildRagPrompt } from './prompt';
+
+describe('buildRagPrompt', () => {
+  it('numbers excerpts 1-based with formatted times and includes the question', () => {
+    const { system, prompt } = buildRagPrompt('What is entropy?', [
+      { text: 'Entropy measures disorder.', start: 65, transcriptId: 't1' },
+      { text: 'It always increases in a closed system.', start: 130, transcriptId: 't1' },
+    ]);
+
+    // Excerpts are numbered and carry a mm:ss timecode from `start`.
+    expect(prompt).toContain('[1] (1:05) Entropy measures disorder.');
+    expect(prompt).toContain('[2] (2:10) It always increases in a closed system.');
+    // Question is present.
+    expect(prompt).toContain('Question: What is entropy?');
+    // System prompt enforces grounding + citations.
+    expect(system).toMatch(/only/i);
+    expect(system).toMatch(/\[1\]|square brackets/i);
+  });
+
+  it('trims the question', () => {
+    const { prompt } = buildRagPrompt('  hello?  ', [
+      { text: 'hi', start: 0, transcriptId: 't1' },
+    ]);
+    expect(prompt).toContain('Question: hello?');
+    expect(prompt).not.toContain('Question:   hello?');
+  });
+
+  it('is deterministic for identical inputs', () => {
+    const snips = [{ text: 'a', start: 5, transcriptId: 't1' }];
+    const a = buildRagPrompt('q', snips);
+    const b = buildRagPrompt('q', snips);
+    expect(a).toEqual(b);
+  });
+
+  it('yields a refusal-style prompt with no snippets (no excerpts -> admit not found)', () => {
+    const { system, prompt } = buildRagPrompt('Anything?', []);
+    expect(prompt).toContain('(no lecture excerpts were found)');
+    expect(prompt).toContain('Question: Anything?');
+    // Still must not contain a fabricated [1] excerpt.
+    expect(prompt).not.toMatch(/^\[1\]/m);
+    // The system rule that drives the refusal is present.
+    expect(system).toMatch(/could not find it in the lectures/i);
+  });
+});

src/lib/generation/prompt.ts

@@ -0,0 +1,56 @@
+// PURE, deterministic RAG prompt builder for the optional "ask your lectures"
+// generative layer (Phase 5). No I/O, no platform deps — just string assembly,
+// so it is fully unit-testable and produces byte-identical output for identical
+// inputs.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+import { formatClock } from '../format';
+
+/** A retrieved lecture snippet to ground the answer in. */
+export interface PromptSnippet {
+  text: string;
+  /** Snippet start, in seconds (rendered as mm:ss in the prompt). */
+  start: number;
+  transcriptId: string;
+}
+
+/**
+ * Build the system + user messages for a grounded, cited RAG answer.
+ *
+ * The system prompt forces the model to answer ONLY from the supplied excerpts,
+ * to admit when the excerpts don't cover the question (no hallucinated answers),
+ * to stay concise, and to cite excerpt numbers like [1], [2].
+ *
+ * The user prompt lists the excerpts as `[i] (mm:ss) text` (1-based), then the
+ * question. With zero snippets the excerpt block is replaced by an explicit
+ * "(no lecture excerpts were found)" marker, which — combined with the grounding
+ * rule — steers the model to a refusal rather than an invented answer.
+ *
+ * Deterministic: same inputs -> same output.
+ */
+export function buildRagPrompt(
+  question: string,
+  snippets: PromptSnippet[],
+): { system: string; prompt: string } {
+  const system = [
+    'You are a study assistant that answers questions about a student\'s lecture recordings.',
+    'Answer ONLY using the numbered lecture excerpts provided below.',
+    'If the excerpts do not contain the answer, say you could not find it in the lectures — do not use outside knowledge and do not guess.',
+    'Be concise.',
+    'Cite the excerpts you used by their number in square brackets, like [1] or [2][3].',
+  ].join(' ');
+
+  const q = question.trim();
+
+  const excerptBlock =
+    snippets.length === 0
+      ? '(no lecture excerpts were found)'
+      : snippets
+          .map((s, i) => `[${i + 1}] (${formatClock(s.start)}) ${s.text.trim()}`)
+          .join('\n');
+
+  const prompt = `Lecture excerpts:\n${excerptBlock}\n\nQuestion: ${q}`;
+
+  return { system, prompt };
+}

src/lib/generation/rag.ts

@@ -0,0 +1,91 @@
+// RAG orchestrator for the optional "ask your lectures" feature (Phase 5).
+//
+// Flow: retrieve the most relevant lecture moments with the existing on-device
+// semantic search, then (only if a generation engine is available) ask the model
+// to write a grounded, cited answer from those moments. The retrieved moments are
+// ALWAYS returned as citations so the UI can show them verbatim.
+//
+// Hard rules honoured here:
+//  - Opt-in + gated: if no engine is available we return the search hits as a
+//    "search-only" fallback (generated:false) — never a fake/ungrounded answer.
+//  - Privacy: we only ever pass the question + retrieved snippets to the engine
+//    (the cloud path is OpenAI-compatible BYO-key); raw audio/transcripts never
+//    leave the device.
+//  - Defensive: any generation failure degrades to the search-only fallback so
+//    the user still gets the moments.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+import { searchLectures } from '../search/search';
+import type { GenerationEngine, RagAnswer, RagCitation } from './engine';
+import { buildRagPrompt } from './prompt';
+
+/** How many lecture moments to retrieve and feed the model. */
+const RAG_LIMIT = 6;
+
+const EMPTY: RagAnswer = {
+  answer: '',
+  citations: [],
+  grounded: false,
+  generated: false,
+};
+
+/**
+ * Answer `question` from the user's lectures, grounded in retrieved moments.
+ *
+ * - Empty/whitespace question -> empty, ungrounded result.
+ * - No relevant moments retrieved -> empty, ungrounded result (we refuse to
+ *   answer with nothing to cite).
+ * - Engine unavailable / generation fails -> the retrieved moments as a
+ *   search-only fallback (grounded:true, generated:false).
+ * - Otherwise -> the model's grounded, cited answer (generated:true).
+ */
+export async function askLectures(
+  question: string,
+  engine: GenerationEngine,
+  opts?: { courseId?: string | null; signal?: AbortSignal },
+): Promise<RagAnswer> {
+  const q = question.trim();
+  if (q.length === 0) return EMPTY;
+
+  const hits = await searchLectures(q, {
+    courseId: opts?.courseId,
+    limit: RAG_LIMIT,
+  });
+  if (hits.length === 0) return EMPTY;
+
+  // The retrieved moments are shown verbatim regardless of whether we generate.
+  const citations: RagCitation[] = hits.map((h) => ({
+    transcriptId: h.transcriptId,
+    segmentId: h.segmentId,
+    start: h.start,
+    text: h.text,
+    score: h.score,
+  }));
+
+  // Search-only fallback: no engine here right now (no WebGPU / no key set).
+  if (!(await engine.isAvailable())) {
+    return { answer: '', citations, grounded: true, generated: false };
+  }
+
+  try {
+    if (!engine.isLoaded()) await engine.loadModel();
+    const { system, prompt } = buildRagPrompt(
+      q,
+      hits.map((h) => ({
+        text: h.text,
+        start: h.start,
+        transcriptId: h.transcriptId,
+      })),
+    );
+    const answer = await engine.generate(prompt, {
+      system,
+      maxTokens: 512,
+      signal: opts?.signal,
+    });
+    return { answer, citations, grounded: true, generated: true };
+  } catch {
+    // Any failure (model load, generation, abort) -> still give the moments.
+    return { answer: '', citations, grounded: true, generated: false };
+  }
+}

src/lib/learn/flashcards.test.ts

@@ -0,0 +1,46 @@
+import { describe, it, expect } from 'vitest';
+import { cardsFromGlossary } from './flashcards';
+import type { GlossaryEntry } from './glossary';
+
+describe('cardsFromGlossary', () => {
+  it('maps one card per entry and carries provenance', () => {
+    const entries: GlossaryEntry[] = [
+      {
+        term: 'Photosynthesis',
+        definition: 'A process used by plants to make food.',
+        start: 7,
+        segmentId: 'segX',
+      },
+    ];
+    const cards = cardsFromGlossary(entries);
+    expect(cards.length).toBe(1);
+    expect(cards[0]!.back).toBe('A process used by plants to make food.');
+    expect(cards[0]!.start).toBe(7);
+    expect(cards[0]!.segmentId).toBe('segX');
+  });
+
+  it('uses a "What is X?" front when the term is absent from the definition', () => {
+    const entries: GlossaryEntry[] = [
+      { term: 'Photosynthesis', definition: 'A process used by plants.', start: 0 },
+    ];
+    const [card] = cardsFromGlossary(entries);
+    expect(card!.front).toBe('What is Photosynthesis?');
+  });
+
+  it('makes a cloze front when the term appears in the definition', () => {
+    const entries: GlossaryEntry[] = [
+      {
+        term: 'Recursion',
+        definition: 'Recursion is when a function calls itself.',
+        start: 0,
+      },
+    ];
+    const [card] = cardsFromGlossary(entries);
+    expect(card!.front).toBe('_____ is when a function calls itself.');
+    expect(card!.front).not.toContain('Recursion');
+  });
+
+  it('returns [] for no entries', () => {
+    expect(cardsFromGlossary([])).toEqual([]);
+  });
+});

src/lib/learn/flashcards.ts

@@ -0,0 +1,51 @@
+// Turn glossary entries into flashcard seeds. Deterministic, no model.
+// A seed is a UI-friendly front/back pair carrying its source provenance; the
+// repo later wraps it into a FlashcardDraft + SRS state.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+import type { GlossaryEntry } from './glossary';
+
+/** A front/back pair for a flashcard, with optional source anchor. */
+export interface CardSeed {
+  front: string;
+  back: string;
+  /** Source segment id, for click-to-seek. */
+  segmentId?: string;
+  /** Source start time (seconds). */
+  start?: number;
+}
+
+/**
+ * Build one {@link CardSeed} per glossary entry.
+ *
+ * Front: if the term appears verbatim (case-insensitively) inside the
+ * definition, we make a cloze deletion — blanking the term in the definition so
+ * the learner recalls it in context. Otherwise we fall back to the plain
+ * "What is {term}?" prompt. Back is always the full definition.
+ */
+export function cardsFromGlossary(entries: GlossaryEntry[]): CardSeed[] {
+  return entries.map((entry) => {
+    const front = makeFront(entry.term, entry.definition);
+    return {
+      front,
+      back: entry.definition,
+      segmentId: entry.segmentId,
+      start: entry.start,
+    };
+  });
+}
+
+/** "_____" cloze if the term occurs in the definition, else a Q prompt. */
+function makeFront(term: string, definition: string): string {
+  const re = new RegExp(escapeRegExp(term), 'i');
+  if (re.test(definition)) {
+    return definition.replace(re, '_____');
+  }
+  return `What is ${term}?`;
+}
+
+/** Escape a string for safe use inside a RegExp. */
+function escapeRegExp(s: string): string {
+  return s.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+}

src/lib/learn/glossary.test.ts

@@ -0,0 +1,64 @@
+import { describe, it, expect } from 'vitest';
+import { glossary } from './glossary';
+import type { Segment } from '../types';
+
+const seg = (start: number, text: string, id?: string): Segment => ({
+  id,
+  start,
+  end: start + 1,
+  text,
+});
+
+describe('glossary', () => {
+  it('returns [] for no segments', () => {
+    expect(glossary([])).toEqual([]);
+  });
+
+  it('captures an "X is Y" definition pattern', () => {
+    const segs = [
+      seg(3, 'Backpropagation is an algorithm for training neural networks.', 's1'),
+    ];
+    const out = glossary(segs);
+    const entry = out.find((e) => e.term.toLowerCase() === 'backpropagation');
+    expect(entry).toBeDefined();
+    expect(entry!.definition).toBe(
+      'Backpropagation is an algorithm for training neural networks.',
+    );
+    expect(entry!.start).toBe(3);
+    expect(entry!.segmentId).toBe('s1');
+  });
+
+  it('recognizes means / refers to / is defined as', () => {
+    const segs = [
+      seg(0, 'Entropy means the amount of disorder in a system.'),
+      seg(1, 'A token refers to a unit of text.'),
+      seg(2, 'Latency is defined as the delay before a transfer begins.'),
+    ];
+    const terms = glossary(segs).map((e) => e.term.toLowerCase());
+    expect(terms).toContain('entropy');
+    expect(terms).toContain('a token');
+    expect(terms).toContain('latency');
+  });
+
+  it('dedupes by lowercased term, preferring the explicit definition', () => {
+    const segs = [
+      seg(0, 'Recursion is fun. Recursion appears again here as Recursion.'),
+      seg(1, 'Recursion is when a function calls itself.', 'def'),
+    ];
+    const out = glossary(segs);
+    const recursion = out.filter((e) => e.term.toLowerCase() === 'recursion');
+    expect(recursion.length).toBe(1);
+    expect(recursion[0]!.definition).toBe(
+      'Recursion is when a function calls itself.',
+    );
+  });
+
+  it('respects the max cap', () => {
+    const segs = [
+      seg(0, 'Apple is a fruit.'),
+      seg(1, 'Banana is a fruit.'),
+      seg(2, 'Cherry is a fruit.'),
+    ];
+    expect(glossary(segs, { max: 2 }).length).toBeLessThanOrEqual(2);
+  });
+});

src/lib/learn/glossary.ts

@@ -0,0 +1,185 @@
+// Deterministic glossary extraction: surface candidate terms + definitions from
+// a lecture transcript with NO model. Two complementary strategies:
+//   (a) explicit definition patterns: "X is/are/means/refers to/is defined as Y"
+//   (b) frequent Capitalized phrases / frequent content nouns as a fallback,
+//       using their first containing sentence as a stand-in definition.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+import type { Segment } from '../types';
+import { splitSentences, tokenizeWords, STOPWORDS } from './tokenize';
+
+/** One glossary row: a term, a definition, and where it came from. */
+export interface GlossaryEntry {
+  term: string;
+  definition: string;
+  /** Start time (seconds) of the source segment. */
+  start: number;
+  /** Source segment id, when the segment had one. */
+  segmentId?: string;
+}
+
+/** Internal accumulator while collecting candidates. */
+interface Candidate extends GlossaryEntry {
+  /** How "important"/frequent this candidate is, for ranking. */
+  frequency: number;
+  /** Whether it came from an explicit definition pattern (preferred). */
+  explicit: boolean;
+  /** Original discovery order, for stable tie-breaking. */
+  order: number;
+}
+
+// "X is/are/means/refers to/is defined as Y" — capture the subject (term) and
+// keep the whole sentence as the definition. The subject is the run of words
+// before the connective; we cap its length so we don't swallow half a clause.
+const DEFINITION_RE =
+  /^(.{2,60}?)\s+(?:is|are|means|refers? to|is defined as)\s+(.+)$/i;
+
+// A run of Capitalized words (proper-noun-ish phrase), e.g. "Hidden Markov Model".
+const CAPITALIZED_PHRASE_RE = /\b([A-Z][a-z0-9]+(?:\s+[A-Z][a-z0-9]+)*)\b/g;
+
+/**
+ * Build a glossary of at most `max` entries from `segments`.
+ *
+ * Candidates are deduped by lowercased term (explicit definitions win over
+ * fallbacks; otherwise the first/most-frequent occurrence is kept), then sorted
+ * by frequency descending (explicit entries ranked ahead on ties), and capped.
+ *
+ * Empty input -> [].
+ */
+export function glossary(
+  segments: Segment[],
+  opts?: { max?: number },
+): GlossaryEntry[] {
+  const max = opts?.max ?? 12;
+  if (segments.length === 0 || max <= 0) return [];
+
+  // Document-wide content-word frequencies, used to rank fallback candidates.
+  const wordFreq = new Map<string, number>();
+  for (const seg of segments) {
+    for (const tok of tokenizeWords(seg.text)) {
+      wordFreq.set(tok, (wordFreq.get(tok) ?? 0) + 1);
+    }
+  }
+
+  // Frequency of each Capitalized phrase (by lowercased key) across the doc.
+  const phraseFreq = new Map<string, number>();
+  for (const seg of segments) {
+    for (const m of seg.text.matchAll(CAPITALIZED_PHRASE_RE)) {
+      const phrase = m[1];
+      if (!phrase) continue;
+      const key = phrase.toLowerCase();
+      phraseFreq.set(key, (phraseFreq.get(key) ?? 0) + 1);
+    }
+  }
+
+  // Dedupe by lowercased term. Map preserves first-seen insertion order.
+  const byTerm = new Map<string, Candidate>();
+  let order = 0;
+
+  const add = (cand: Omit<Candidate, 'order'>) => {
+    const key = cand.term.toLowerCase().trim();
+    if (key.length === 0) return;
+    const existing = byTerm.get(key);
+    if (!existing) {
+      byTerm.set(key, { ...cand, order: order++ });
+      return;
+    }
+    // An explicit definition supersedes a fallback for the same term. When both
+    // are explicit, keep the more informative (longer) definition; this lets a
+    // substantive "X is when ..." win over a throwaway "X is fun." mention.
+    const upgrade =
+      (cand.explicit && !existing.explicit) ||
+      (cand.explicit === existing.explicit &&
+        cand.definition.length > existing.definition.length);
+    if (upgrade) {
+      byTerm.set(key, { ...cand, order: existing.order });
+    }
+  };
+
+  // Pass 1: explicit "X is Y" definitions (highest quality).
+  for (const seg of segments) {
+    for (const sentence of splitSentences(seg.text)) {
+      const m = sentence.match(DEFINITION_RE);
+      if (!m || !m[1]) continue;
+      const term = m[1].trim();
+      // Skip junk subjects: a term with no content words (pure stopwords) OR one
+      // that's too short. (Was `&&`, which only skipped when BOTH held, so an
+      // all-stopword term like "there" still became a glossary entry.)
+      const termWords = tokenizeWords(term);
+      if (termWords.length === 0 || term.length < 3) continue;
+      add({
+        term,
+        definition: sentence,
+        start: seg.start,
+        segmentId: seg.id,
+        // Frequency from the phrase/word tables, boosted so explicit wins.
+        frequency:
+          (phraseFreq.get(term.toLowerCase()) ?? 0) +
+          (wordFreq.get(term.toLowerCase()) ?? 0) +
+          1,
+        explicit: true,
+      });
+    }
+  }
+
+  // Pass 2a: frequent multi-word Capitalized phrases as fallback candidates.
+  for (const seg of segments) {
+    const sentences = splitSentences(seg.text);
+    for (const [key, freq] of phraseFreq) {
+      if (freq < 1) continue;
+      // Only multi-word phrases here; single Capitalized words are noisy.
+      if (!key.includes(' ')) continue;
+      // First sentence in this segment that contains the phrase.
+      const containing = sentences.find((s) => s.toLowerCase().includes(key));
+      if (!containing) continue;
+      // Reconstruct a Title-Cased display term from the key.
+      const term = key.replace(/\b\w/g, (c) => c.toUpperCase());
+      add({
+        term,
+        definition: containing,
+        start: seg.start,
+        segmentId: seg.id,
+        frequency: freq,
+        explicit: false,
+      });
+    }
+  }
+
+  // Pass 2b: frequent single content nouns as a last-resort fallback.
+  for (const seg of segments) {
+    const sentences = splitSentences(seg.text);
+    for (const tok of tokenizeWords(seg.text)) {
+      if (STOPWORDS.has(tok)) continue;
+      const freq = wordFreq.get(tok) ?? 0;
+      if (freq < 2) continue; // require some recurrence to qualify
+      const containing = sentences.find((s) =>
+        s.toLowerCase().includes(tok),
+      );
+      if (!containing) continue;
+      add({
+        term: tok,
+        definition: containing,
+        start: seg.start,
+        segmentId: seg.id,
+        frequency: freq,
+        explicit: false,
+      });
+    }
+  }
+
+  // Rank: explicit first, then frequency desc, then discovery order (stable).
+  const ranked = [...byTerm.values()].sort(
+    (a, b) =>
+      Number(b.explicit) - Number(a.explicit) ||
+      b.frequency - a.frequency ||
+      a.order - b.order,
+  );
+
+  return ranked.slice(0, max).map(({ term, definition, start, segmentId }) => ({
+    term,
+    definition,
+    start,
+    segmentId,
+  }));
+}

src/lib/learn/index.ts

@@ -0,0 +1,15 @@
+// Public surface of the deterministic study helpers (Phase 3). All pure: no
+// React/Expo/Node-builtins, so the UI and tests can import freely.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+export { splitSentences, tokenizeWords, STOPWORDS } from './tokenize';
+export { summarize } from './summary';
+export type { SourcedSentence } from './summary';
+export { glossary } from './glossary';
+export type { GlossaryEntry } from './glossary';
+export { cardsFromGlossary } from './flashcards';
+export type { CardSeed } from './flashcards';
+export { initialSrs, review } from './srs';
+export { generateQuiz } from './quiz';
+export type { QuizQuestion } from './quiz';

src/lib/learn/quiz.test.ts

@@ -0,0 +1,66 @@
+import { describe, it, expect } from 'vitest';
+import { generateQuiz } from './quiz';
+import type { GlossaryEntry } from './glossary';
+
+const entry = (term: string, definition: string, start = 0): GlossaryEntry => ({
+  term,
+  definition,
+  start,
+  segmentId: `seg-${term}`,
+});
+
+const FOUR: GlossaryEntry[] = [
+  entry('Alpha', 'the first letter'),
+  entry('Beta', 'the second letter'),
+  entry('Gamma', 'the third letter'),
+  entry('Delta', 'the fourth letter'),
+];
+
+describe('generateQuiz', () => {
+  it('returns [] when there are fewer than 4 terms', () => {
+    expect(generateQuiz(FOUR.slice(0, 3))).toEqual([]);
+    expect(generateQuiz([])).toEqual([]);
+  });
+
+  it('returns [] when count <= 0', () => {
+    expect(generateQuiz(FOUR, { count: 0 })).toEqual([]);
+  });
+
+  it('each question has 4 options with the answer at answerIndex', () => {
+    const quiz = generateQuiz(FOUR);
+    expect(quiz.length).toBe(4);
+    for (let i = 0; i < quiz.length; i++) {
+      const q = quiz[i]!;
+      const src = FOUR[i]!;
+      expect(q.options.length).toBe(4);
+      // The option at answerIndex is the correct term for that question.
+      expect(q.options[q.answerIndex]).toBe(src.term);
+      // Question wraps the correct definition.
+      expect(q.question).toBe(`Which term means: "${src.definition}"?`);
+    }
+  });
+
+  it('uses deterministic answer placement at index i % 4', () => {
+    const quiz = generateQuiz(FOUR);
+    expect(quiz.map((q) => q.answerIndex)).toEqual([0, 1, 2, 3]);
+  });
+
+  it('options are distinct and include 3 distractors from other entries', () => {
+    const quiz = generateQuiz(FOUR);
+    for (const q of quiz) {
+      const unique = new Set(q.options.map((o) => o.toLowerCase()));
+      expect(unique.size).toBe(4);
+    }
+  });
+
+  it('carries start and segmentId from the correct entry', () => {
+    const quiz = generateQuiz(FOUR);
+    expect(quiz[0]!.segmentId).toBe('seg-Alpha');
+    expect(quiz[0]!.start).toBe(0);
+  });
+
+  it('respects the count cap', () => {
+    const quiz = generateQuiz(FOUR, { count: 2 });
+    expect(quiz.length).toBe(2);
+  });
+});

src/lib/learn/quiz.ts

@@ -0,0 +1,88 @@
+// Deterministic multiple-choice quiz generation from glossary entries. No model,
+// NO randomness — given the same entries, the same quiz is produced every time
+// (important for reproducible tests and stable UI across re-renders).
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+import type { GlossaryEntry } from './glossary';
+
+/** One MCQ: a definition prompt, term options, and the correct option index. */
+export interface QuizQuestion {
+  question: string;
+  options: string[];
+  answerIndex: number;
+  /** Source start time (seconds). */
+  start?: number;
+  /** Source segment id. */
+  segmentId?: string;
+}
+
+const OPTION_COUNT = 4;
+
+/**
+ * Generate up to `count` questions from `entries`.
+ *
+ * Requires at least {@link OPTION_COUNT} (4) distinct terms so every question
+ * can have one correct answer + 3 distinct distractors; with fewer entries we
+ * return [].
+ *
+ * Determinism: for question `i`, the correct answer is placed at index
+ * `i % OPTION_COUNT`, and the remaining slots are filled with distractor terms
+ * drawn from the OTHER entries by rotating through the list (offset by `i`), so
+ * the layout is a pure function of the input order.
+ */
+export function generateQuiz(
+  entries: GlossaryEntry[],
+  opts?: { count?: number },
+): QuizQuestion[] {
+  const count = opts?.count ?? 8;
+  if (entries.length < OPTION_COUNT || count <= 0) return [];
+
+  const questions: QuizQuestion[] = [];
+  const n = entries.length;
+  const limit = Math.min(count, n);
+
+  for (let i = 0; i < limit; i++) {
+    const correct = entries[i];
+    if (!correct) continue; // unreachable (i < limit <= n), satisfies the checker
+
+    // Collect 3 distinct distractor terms from other entries, rotating the
+    // start point by `i` so different questions get different wrong answers.
+    const distractors: string[] = [];
+    const seen = new Set<string>([correct.term.toLowerCase()]);
+    let j = 1;
+    while (distractors.length < OPTION_COUNT - 1 && j <= n) {
+      const cand = entries[(i + j) % n];
+      j++;
+      if (!cand) continue;
+      const key = cand.term.toLowerCase();
+      if (!seen.has(key)) {
+        seen.add(key);
+        distractors.push(cand.term);
+      }
+    }
+
+    // Defensive: if duplicate terms left us short, skip this question rather
+    // than emit one with fewer than 4 options. (glossary() dedupes, so this is
+    // effectively unreachable for real input.)
+    if (distractors.length < OPTION_COUNT - 1) continue;
+
+    // Place the answer at a deterministic index; fill the rest with distractors.
+    const answerIndex = i % OPTION_COUNT;
+    const options: string[] = [];
+    let d = 0;
+    for (let slot = 0; slot < OPTION_COUNT; slot++) {
+      options.push(slot === answerIndex ? correct.term : (distractors[d++] as string));
+    }
+
+    questions.push({
+      question: `Which term means: "${correct.definition}"?`,
+      options,
+      answerIndex,
+      start: correct.start,
+      segmentId: correct.segmentId,
+    });
+  }
+
+  return questions;
+}

src/lib/learn/srs.test.ts

@@ -0,0 +1,95 @@
+import { describe, it, expect } from 'vitest';
+import { initialSrs, review } from './srs';
+
+const MS_PER_DAY = 86_400_000;
+const NOW = 1_700_000_000_000;
+
+describe('initialSrs', () => {
+  it('creates a neutral, immediately-due card', () => {
+    const s = initialSrs(NOW);
+    expect(s).toEqual({
+      ease: 2.5,
+      intervalDays: 0,
+      reps: 0,
+      lapses: 0,
+      due: NOW,
+    });
+  });
+});
+
+describe('review', () => {
+  it('does not mutate the input state', () => {
+    const s = initialSrs(NOW);
+    const frozen = { ...s };
+    review(s, 2, NOW);
+    expect(s).toEqual(frozen);
+  });
+
+  it('produces monotonically increasing intervals on repeated "good"', () => {
+    let s = initialSrs(NOW);
+    const intervals: number[] = [];
+    let t = NOW;
+    for (let i = 0; i < 5; i++) {
+      s = review(s, 2, t);
+      intervals.push(s.intervalDays);
+      t += s.intervalDays * MS_PER_DAY;
+    }
+    // 1 day, 6 days, then *ease each time -> strictly increasing.
+    for (let i = 1; i < intervals.length; i++) {
+      expect(intervals[i]!).toBeGreaterThan(intervals[i - 1]!);
+    }
+    expect(intervals[0]).toBe(1);
+    expect(intervals[1]).toBe(6);
+  });
+
+  it('increments reps on passing grades and resets on "again"', () => {
+    let s = initialSrs(NOW);
+    s = review(s, 2, NOW); // good
+    s = review(s, 2, NOW); // good
+    expect(s.reps).toBe(2);
+    s = review(s, 0, NOW); // again
+    expect(s.reps).toBe(0);
+    expect(s.lapses).toBe(1);
+  });
+
+  it('floors ease at 1.3 even after many "again" reviews', () => {
+    let s = initialSrs(NOW);
+    for (let i = 0; i < 20; i++) s = review(s, 0, NOW);
+    expect(s.ease).toBe(1.3);
+  });
+
+  it('lowers ease on hard, raises it on easy', () => {
+    const good = review(initialSrs(NOW), 2, NOW);
+    expect(good.ease).toBe(2.5);
+    const hard = review(initialSrs(NOW), 1, NOW);
+    expect(hard.ease).toBeCloseTo(2.35, 5);
+    const easy = review(initialSrs(NOW), 3, NOW);
+    expect(easy.ease).toBeCloseTo(2.65, 5);
+  });
+
+  it('computes due = now + intervalDays * 86400000 and sets lastReviewed', () => {
+    const s = review(initialSrs(NOW), 2, NOW);
+    expect(s.intervalDays).toBe(1);
+    expect(s.due).toBe(NOW + 1 * MS_PER_DAY);
+    expect(s.lastReviewed).toBe(NOW);
+  });
+
+  it('gives "easy" a longer interval than "good" at the same step', () => {
+    const good = review(initialSrs(NOW), 2, NOW);
+    const easy = review(initialSrs(NOW), 3, NOW);
+    expect(easy.intervalDays).toBeGreaterThan(good.intervalDays);
+  });
+
+  it('keeps "hard" shorter than "good" for a reviewed card (reps >= 2)', () => {
+    // Bring a card to reps=2 with two Goods (interval 6, ease 2.5).
+    let s = initialSrs(NOW);
+    s = review(s, 2, NOW);
+    s = review(s, 2, NOW);
+    expect(s.reps).toBe(2);
+    const hard = review(s, 1, NOW);
+    const good = review(s, 2, NOW);
+    expect(hard.intervalDays).toBeLessThan(good.intervalDays);
+    // And Hard still grows the interval (it's a pass, not a lapse).
+    expect(hard.intervalDays).toBeGreaterThan(s.intervalDays);
+  });
+});

src/lib/learn/srs.ts

@@ -0,0 +1,124 @@
+// Spaced repetition scheduling — a faithful SM-2 variant adapted to a 4-button
+// grading UI (Again / Hard / Good / Easy). Pure & deterministic.
+//
+// The classic SM-2 algorithm grades recall on 0..5 and recomputes an ease
+// factor (EF) plus an interval. We collapse the UI to four buttons and map them
+// to behaviors that match learner expectations on Anki-style apps:
+//
+//   grade 0 = Again : failed recall. Reset the rep streak, count a lapse, drop
+//                     ease by 0.2, and re-show soon (a short sub-day interval).
+//   grade 1 = Hard  : recalled with difficulty. Grow the interval gently (x1.2)
+//                     and drop ease by 0.15.
+//   grade 2 = Good  : normal recall. Use the SM-2 schedule:
+//                     rep 1 -> 1 day, rep 2 -> 6 days, then interval *= ease.
+//   grade 3 = Easy  : effortless recall. Like Good but with an easy bonus
+//                     (~x1.3) and ease bumped up by 0.15.
+//
+// Invariants:
+//   - ease is floored at 1.3 (SM-2's classic minimum).
+//   - reps increments only on a passing grade (>= 1); Again resets it to 0.
+//   - due = now + intervalDays * MS_PER_DAY.
+//   - lastReviewed is set to `now` on every review.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+import type { SrsState } from '../db/schema';
+
+const MS_PER_DAY = 86_400_000;
+const MIN_EASE = 1.3;
+/** Sub-day interval (in days) used when a card is failed and must re-show soon. */
+const AGAIN_INTERVAL_DAYS = 1 / 1440; // ~1 minute, expressed in days
+const HARD_MULTIPLIER = 1.2;
+const EASY_BONUS = 1.3;
+
+/** A brand-new card: due immediately, neutral ease, nothing learned yet. */
+export function initialSrs(now: number): SrsState {
+  return {
+    ease: 2.5,
+    intervalDays: 0,
+    reps: 0,
+    lapses: 0,
+    due: now,
+  };
+}
+
+/**
+ * Apply a review `grade` to `srs` at time `now`, returning the next state.
+ * Does not mutate the input.
+ */
+export function review(
+  srs: SrsState,
+  grade: 0 | 1 | 2 | 3,
+  now: number,
+): SrsState {
+  let { ease, intervalDays, reps, lapses } = srs;
+
+  if (grade === 0) {
+    // Again: the card was forgotten.
+    ease = floorEase(ease - 0.2);
+    reps = 0;
+    lapses += 1;
+    intervalDays = AGAIN_INTERVAL_DAYS;
+  } else if (grade === 1) {
+    // Hard: passed with difficulty. Lower ease and grow the interval only
+    // GENTLY (x1.2) off the PREVIOUS interval — crucially WITHOUT compounding
+    // the ease factor. Multiplying by ease as well would push Hard past Good for
+    // reviewed cards (e.g. 2.35*1.2 = 2.82 > 2.5), the opposite of what a
+    // struggling learner expects.
+    ease = floorEase(ease - 0.15);
+    reps += 1;
+    intervalDays = hardInterval(srs);
+  } else if (grade === 2) {
+    // Good: standard SM-2 progression, ease unchanged.
+    reps += 1;
+    intervalDays = nextInterval(srs, ease, 1);
+  } else {
+    // Easy: like Good with an easy bonus and an ease bump.
+    ease = floorEase(ease + 0.15);
+    reps += 1;
+    intervalDays = nextInterval(srs, ease, EASY_BONUS);
+  }
+
+  return {
+    ease,
+    intervalDays,
+    reps,
+    lapses,
+    due: now + intervalDays * MS_PER_DAY,
+    lastReviewed: now,
+  };
+}
+
+/**
+ * SM-2 interval progression for a passing grade.
+ *  - first successful rep (the card had reps 0) -> 1 day
+ *  - second successful rep (had reps 1)         -> 6 days
+ *  - thereafter                                 -> previous interval * ease
+ * The result is scaled by `multiplier` (Hard < 1-ish via lower ease, Easy bonus).
+ */
+function nextInterval(prev: SrsState, ease: number, multiplier: number): number {
+  let base: number;
+  if (prev.reps <= 0) {
+    base = 1;
+  } else if (prev.reps === 1) {
+    base = 6;
+  } else {
+    base = prev.intervalDays * ease;
+  }
+  return base * multiplier;
+}
+
+/**
+ * Interval for a "Hard" pass: a gentle x1.2 growth off the previous interval,
+ * never compounded by ease, so Hard stays strictly shorter than Good (which
+ * multiplies by ease) at every stage. A brand-new/unseen card gets a 1-day step.
+ */
+function hardInterval(prev: SrsState): number {
+  if (prev.reps <= 0 || prev.intervalDays <= 0) return 1;
+  return prev.intervalDays * HARD_MULTIPLIER;
+}
+
+/** Clamp an ease factor to SM-2's minimum of 1.3. */
+function floorEase(ease: number): number {
+  return ease < MIN_EASE ? MIN_EASE : ease;
+}

src/lib/learn/summary.test.ts

@@ -0,0 +1,57 @@
+import { describe, it, expect } from 'vitest';
+import { summarize } from './summary';
+import type { Segment } from '../types';
+
+const seg = (start: number, text: string, id?: string): Segment => ({
+  id,
+  start,
+  end: start + 1,
+  text,
+});
+
+describe('summarize', () => {
+  it('returns [] for no segments', () => {
+    expect(summarize([])).toEqual([]);
+  });
+
+  it('returns [] when maxSentences <= 0', () => {
+    expect(summarize([seg(0, 'Neural networks learn.')], { maxSentences: 0 })).toEqual(
+      [],
+    );
+  });
+
+  it('picks the top-N highest-frequency sentences', () => {
+    // "neural" + "networks" dominate; the off-topic sentence should be dropped.
+    const segs = [
+      seg(0, 'Neural networks process data.'),
+      seg(1, 'Neural networks learn from neural networks.'),
+      seg(2, 'The weather today is sunny.'),
+    ];
+    const out = summarize(segs, { maxSentences: 2 });
+    expect(out.length).toBe(2);
+    const texts = out.map((s) => s.text);
+    expect(texts).not.toContain('The weather today is sunny.');
+  });
+
+  it('returns sentences in chronological order with the right start + id', () => {
+    const segs = [
+      seg(10, 'Gradient descent optimizes the gradient.', 'segB'),
+      seg(2, 'Gradient descent uses the gradient slope.', 'segA'),
+    ];
+    const out = summarize(segs, { maxSentences: 2 });
+    // Both kept; chronological order means start 2 comes before start 10.
+    expect(out.map((s) => s.start)).toEqual([2, 10]);
+    expect(out[0]!.segmentId).toBe('segA');
+    expect(out[1]!.segmentId).toBe('segB');
+  });
+
+  it('flattens multi-sentence segments and tags them with the segment start', () => {
+    const segs = [seg(5, 'Alpha beta gamma. Alpha beta delta.', 'multi')];
+    const out = summarize(segs, { maxSentences: 5 });
+    expect(out.length).toBe(2);
+    for (const s of out) {
+      expect(s.start).toBe(5);
+      expect(s.segmentId).toBe('multi');
+    }
+  });
+});

src/lib/learn/summary.ts

@@ -0,0 +1,90 @@
+// Deterministic extractive summary: pick the most "central" sentences of a
+// lecture by term-frequency scoring, keeping them in chronological order so the
+// summary still reads like a condensed version of the talk. No model.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+import type { Segment } from '../types';
+import { splitSentences, tokenizeWords } from './tokenize';
+
+/** A summary sentence tagged with where it came from (for click-to-seek). */
+export interface SourcedSentence {
+  text: string;
+  /** Start time (seconds) of the source segment. */
+  start: number;
+  /** Source segment id, when the segment had one. */
+  segmentId?: string;
+}
+
+/** Internal: a sentence carrying its source + original ordering keys. */
+interface ScoredSentence extends SourcedSentence {
+  /** Original position across the flattened doc, for stable chronological sort. */
+  position: number;
+  /** Term-frequency score, normalized by sentence length. */
+  score: number;
+}
+
+/**
+ * Summarize `segments` into at most `maxSentences` sentences.
+ *
+ * Algorithm (fully deterministic):
+ *  1. Build a document term-frequency table over `tokenizeWords` of all text.
+ *  2. Flatten every segment into sentences, each tagged with that segment's
+ *     start + id, and a global position index.
+ *  3. Score each sentence as the sum of its content words' document frequencies,
+ *     normalized by the number of content words (so long sentences don't win by
+ *     length alone). Sentences with no content words score 0.
+ *  4. Take the top `maxSentences` by score (ties broken by earlier position),
+ *     then RETURN them re-sorted into chronological order (start, then position).
+ *
+ * Empty input -> [].
+ */
+export function summarize(
+  segments: Segment[],
+  opts?: { maxSentences?: number },
+): SourcedSentence[] {
+  const maxSentences = opts?.maxSentences ?? 5;
+  if (segments.length === 0 || maxSentences <= 0) return [];
+
+  // 1. Document term frequencies.
+  const freq = new Map<string, number>();
+  for (const seg of segments) {
+    for (const tok of tokenizeWords(seg.text)) {
+      freq.set(tok, (freq.get(tok) ?? 0) + 1);
+    }
+  }
+
+  // 2. Flatten into sentences with provenance + position.
+  const sentences: ScoredSentence[] = [];
+  let position = 0;
+  for (const seg of segments) {
+    for (const text of splitSentences(seg.text)) {
+      // 3. Score this sentence.
+      const tokens = tokenizeWords(text);
+      let sum = 0;
+      for (const tok of tokens) sum += freq.get(tok) ?? 0;
+      const score = tokens.length > 0 ? sum / tokens.length : 0;
+
+      sentences.push({
+        text,
+        start: seg.start,
+        segmentId: seg.id,
+        position: position++,
+        score,
+      });
+    }
+  }
+
+  if (sentences.length === 0) return [];
+
+  // 4a. Pick top-N by score, ties broken by earlier position (deterministic).
+  const ranked = [...sentences].sort(
+    (a, b) => b.score - a.score || a.position - b.position,
+  );
+  const picked = ranked.slice(0, maxSentences);
+
+  // 4b. Restore chronological order: by start, then original position.
+  picked.sort((a, b) => a.start - b.start || a.position - b.position);
+
+  return picked.map(({ text, start, segmentId }) => ({ text, start, segmentId }));
+}

src/lib/learn/tokenize.test.ts

@@ -0,0 +1,50 @@
+import { describe, it, expect } from 'vitest';
+import { splitSentences, tokenizeWords, STOPWORDS } from './tokenize';
+
+describe('splitSentences', () => {
+  it('splits on . ? and ! and trims', () => {
+    const out = splitSentences('Hello world. How are you? Great!');
+    expect(out).toEqual(['Hello world.', 'How are you?', 'Great!']);
+  });
+
+  it('splits on newlines too', () => {
+    const out = splitSentences('first line\nsecond line\r\nthird line');
+    expect(out).toEqual(['first line', 'second line', 'third line']);
+  });
+
+  it('drops empty fragments and whitespace-only pieces', () => {
+    expect(splitSentences('   ')).toEqual([]);
+    expect(splitSentences('A.  . B.')).toEqual(['A.', '.', 'B.']);
+    expect(splitSentences('')).toEqual([]);
+  });
+
+  it('keeps the terminating punctuation with its sentence', () => {
+    expect(splitSentences('One. Two.')).toEqual(['One.', 'Two.']);
+  });
+});
+
+describe('tokenizeWords', () => {
+  it('lowercases and splits on non-word runs', () => {
+    expect(tokenizeWords('Neural-Networks rock')).toEqual([
+      'neural',
+      'networks',
+      'rock',
+    ]);
+  });
+
+  it('drops empties, stopwords, and tokens shorter than 3 chars', () => {
+    // "is", "a", "of" are short/stop; "the" is a stopword.
+    const out = tokenizeWords('The cat is on a mat of fur');
+    expect(out).toEqual(['cat', 'mat', 'fur']);
+  });
+
+  it('uses the shared STOPWORD set', () => {
+    expect(STOPWORDS.has('the')).toBe(true);
+    expect(tokenizeWords('the and for')).toEqual([]);
+  });
+
+  it('returns [] for empty / symbol-only input', () => {
+    expect(tokenizeWords('')).toEqual([]);
+    expect(tokenizeWords('!!! ??? ...')).toEqual([]);
+  });
+});

src/lib/learn/tokenize.ts

@@ -0,0 +1,127 @@
+// Text tokenization for the deterministic study helpers (summary, glossary,
+// flashcards, quiz). No model, no I/O, no platform deps — pure & testable.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+/**
+ * A small, language-agnostic-leaning English stopword set. Kept intentionally
+ * compact: enough to stop "the/and/of" from dominating term frequency, without
+ * stripping domain words. Used by both summary scoring and glossary candidate
+ * selection.
+ */
+export const STOPWORDS: ReadonlySet<string> = new Set([
+  'the',
+  'and',
+  'for',
+  'are',
+  'but',
+  'not',
+  'you',
+  'all',
+  'any',
+  'can',
+  'had',
+  'her',
+  'was',
+  'one',
+  'our',
+  'out',
+  'has',
+  'his',
+  'how',
+  'its',
+  'may',
+  'new',
+  'now',
+  'old',
+  'see',
+  'two',
+  'way',
+  'who',
+  'did',
+  'get',
+  'him',
+  'use',
+  'this',
+  'that',
+  'with',
+  'from',
+  'they',
+  'will',
+  'what',
+  'when',
+  'were',
+  'been',
+  'have',
+  'into',
+  'than',
+  'them',
+  'then',
+  'some',
+  'such',
+  'also',
+  'just',
+  'like',
+  'over',
+  'only',
+  'most',
+  'much',
+  'very',
+  'each',
+  'because',
+  'about',
+  'which',
+  'their',
+  'there',
+  'these',
+  'those',
+  'would',
+  'could',
+  'should',
+  'where',
+  'while',
+  'being',
+  'between',
+  'here',
+  'does',
+  'doing',
+  'done',
+  'made',
+  'make',
+  'used',
+  'using',
+  'both',
+  'more',
+  'less',
+]);
+
+/**
+ * Split text into sentences on sentence-ending punctuation (`.`, `?`, `!`) and
+ * newlines. Each result is trimmed; empties are dropped.
+ *
+ * This is intentionally simple (no abbreviation handling) — lecture transcripts
+ * are conversational and rarely contain "Dr." / "e.g." style edge cases, and a
+ * mis-split sentence only marginally affects summary/glossary scoring.
+ */
+export function splitSentences(text: string): string[] {
+  return text
+    // Break after any run of .?! (keep the run with the sentence it ends), and
+    // also break on newlines so transcript line breaks become boundaries.
+    .split(/(?<=[.?!])\s+|[\r\n]+/)
+    .map((s) => s.trim())
+    .filter((s) => s.length > 0);
+}
+
+/**
+ * Tokenize text into "content words": lowercase, split on non-word runs,
+ * dropping empties, stopwords, and very short tokens (length < 3).
+ *
+ * Used for term-frequency scoring; short tokens and stopwords carry little
+ * topical signal and would otherwise dominate the frequency counts.
+ */
+export function tokenizeWords(text: string): string[] {
+  return text
+    .toLowerCase()
+    .split(/\W+/)
+    .filter((t) => t.length >= 3 && !STOPWORDS.has(t));
+}

src/lib/search/lexical.test.ts

@@ -0,0 +1,68 @@
+import { describe, it, expect } from 'vitest';
+import { lexicalRank } from './lexical';
+
+describe('lexicalRank', () => {
+  it('returns [] for an empty query', () => {
+    const rows = [{ id: 'a', text: 'hello world' }];
+    expect(lexicalRank(rows, '')).toEqual([]);
+    expect(lexicalRank(rows, '   ')).toEqual([]);
+    expect(lexicalRank(rows, '!!! ???')).toEqual([]);
+  });
+
+  it('returns [] when no row matches', () => {
+    const rows = [
+      { id: 'a', text: 'the quick brown fox' },
+      { id: 'b', text: 'lazy dog' },
+    ];
+    expect(lexicalRank(rows, 'elephant')).toEqual([]);
+  });
+
+  it('scores by total term-occurrence count', () => {
+    const rows = [
+      { id: 'a', text: 'cat cat cat' },
+      { id: 'b', text: 'cat dog' },
+    ];
+    const out = lexicalRank(rows, 'cat');
+    expect(out).toEqual([
+      { id: 'a', score: 3 },
+      { id: 'b', score: 1 },
+    ]);
+  });
+
+  it('is case-insensitive and tokenizes on non-word chars', () => {
+    const rows = [{ id: 'a', text: 'Neural-Networks, neural networks!' }];
+    // "neural" appears twice, "networks" appears twice => 4.
+    expect(lexicalRank(rows, 'Neural networks')).toEqual([{ id: 'a', score: 4 }]);
+  });
+
+  it('sums distinct query terms and drops zero-score rows', () => {
+    const rows = [
+      { id: 'a', text: 'gradient descent and gradient ascent' }, // gradient x2
+      { id: 'b', text: 'gradient boosting' }, // gradient x1
+      { id: 'c', text: 'random forest' }, // 0 -> dropped
+    ];
+    const out = lexicalRank(rows, 'gradient descent');
+    // a: gradient(2)+descent(1)=3 ; b: gradient(1)=1 ; c dropped
+    expect(out).toEqual([
+      { id: 'a', score: 3 },
+      { id: 'b', score: 1 },
+    ]);
+  });
+
+  it('sorts by score descending', () => {
+    const rows = [
+      { id: 'low', text: 'alpha' },
+      { id: 'high', text: 'alpha alpha alpha' },
+      { id: 'mid', text: 'alpha alpha' },
+    ];
+    expect(lexicalRank(rows, 'alpha').map((r) => r.id)).toEqual(['high', 'mid', 'low']);
+  });
+
+  it('skips rows with empty text', () => {
+    const rows = [
+      { id: 'empty', text: '' },
+      { id: 'a', text: 'match match' },
+    ];
+    expect(lexicalRank(rows, 'match')).toEqual([{ id: 'a', score: 2 }]);
+  });
+});

src/lib/search/lexical.ts

@@ -0,0 +1,44 @@
+// Pure lexical (keyword) ranking. Complements semantic search by catching exact
+// term matches the embedding model might under-weight (names, acronyms, codes).
+// No I/O, no platform deps — fully unit-testable.
+
+/** Lowercase + split on non-word chars, dropping empties. */
+function tokenize(s: string): string[] {
+  return s
+    .toLowerCase()
+    .split(/\W+/)
+    .filter((t) => t.length > 0);
+}
+
+/**
+ * Rank `rows` by how often the query's terms occur in each row's text.
+ *
+ * Scoring: sum, over each query term, of the number of times that term appears
+ * as a token in the row's text. Rows scoring zero are dropped. Result is sorted
+ * by score descending (ties keep input order, since Array.sort is stable).
+ */
+export function lexicalRank(
+  rows: { id: string; text: string }[],
+  query: string,
+): { id: string; score: number }[] {
+  const terms = tokenize(query);
+  if (terms.length === 0) return [];
+
+  const out: { id: string; score: number }[] = [];
+  for (const row of rows) {
+    const tokens = tokenize(row.text);
+    if (tokens.length === 0) continue;
+
+    // Count occurrences per token once, then sum the query terms' counts.
+    const counts = new Map<string, number>();
+    for (const tok of tokens) counts.set(tok, (counts.get(tok) ?? 0) + 1);
+
+    let score = 0;
+    for (const term of terms) score += counts.get(term) ?? 0;
+
+    if (score > 0) out.push({ id: row.id, score });
+  }
+
+  out.sort((a, b) => b.score - a.score);
+  return out;
+}

src/lib/search/rrf.test.ts

@@ -0,0 +1,55 @@
+import { describe, it, expect } from 'vitest';
+import { rrf } from './rrf';
+
+describe('rrf', () => {
+  it('returns [] for no lists / empty lists', () => {
+    expect(rrf([])).toEqual([]);
+    expect(rrf([[], []])).toEqual([]);
+  });
+
+  it('ranks a single list by its existing order', () => {
+    const out = rrf([['a', 'b', 'c']]);
+    expect(out.map((r) => r.id)).toEqual(['a', 'b', 'c']);
+    // rank-1 score = 1/(60+1)
+    expect(out[0]!.score).toBeCloseTo(1 / 61, 10);
+    expect(out[1]!.score).toBeCloseTo(1 / 62, 10);
+  });
+
+  it('sums contributions for ids in multiple lists', () => {
+    // 'a' is rank 1 in list1 and rank 1 in list2 => 2/(k+1).
+    const out = rrf([
+      ['a', 'b'],
+      ['a', 'c'],
+    ]);
+    const a = out.find((r) => r.id === 'a')!;
+    expect(a.score).toBeCloseTo(2 / 61, 10);
+    // 'a' should rank above 'b' and 'c' (each only 1/(k+2)).
+    expect(out[0]!.id).toBe('a');
+  });
+
+  it('an item ranked high in both lists beats one ranked high in only one', () => {
+    const semantic = ['x', 'y', 'z'];
+    const lexical = ['y', 'x', 'w'];
+    const out = rrf([semantic, lexical]);
+    // y: 1/(60+2) + 1/(60+1) ; x: 1/(60+1) + 1/(60+2) -> tie, but both top.
+    expect(new Set([out[0]!.id, out[1]!.id])).toEqual(new Set(['x', 'y']));
+    // z and w each appear once, so they rank lower.
+    expect(out.slice(2).map((r) => r.id).sort()).toEqual(['w', 'z']);
+  });
+
+  it('respects a custom k', () => {
+    const out = rrf([['a']], 0);
+    // rank 1 with k=0 => 1/1 = 1.
+    expect(out[0]!.score).toBeCloseTo(1, 10);
+  });
+
+  it('sorts by fused score descending', () => {
+    const out = rrf([
+      ['a', 'b', 'c'],
+      ['a', 'b', 'c'],
+    ]);
+    expect(out.map((r) => r.id)).toEqual(['a', 'b', 'c']);
+    expect(out[0]!.score).toBeGreaterThan(out[1]!.score);
+    expect(out[1]!.score).toBeGreaterThan(out[2]!.score);
+  });
+});

src/lib/search/rrf.ts

@@ -0,0 +1,30 @@
+// Pure Reciprocal Rank Fusion (RRF). Combines several ranked id lists into one
+// ranking by summing 1/(k + rank) for each id across the lists it appears in
+// (rank is 1-based; k dampens the contribution of low ranks). This is a robust,
+// score-free way to fuse heterogeneous signals (semantic + lexical) — only the
+// ORDER of each input list matters, not its raw scores.
+//
+// No I/O, no platform deps — fully unit-testable.
+
+/**
+ * Fuse ordered id lists into a single ranking, highest fused score first.
+ *
+ * @param lists ordered id lists (rank 1 = first element of each list).
+ * @param k     RRF damping constant (default 60, the canonical value).
+ */
+export function rrf(lists: string[][], k = 60): { id: string; score: number }[] {
+  const scores = new Map<string, number>();
+
+  for (const list of lists) {
+    for (let i = 0; i < list.length; i++) {
+      const id = list[i];
+      if (id === undefined) continue;
+      const rank = i + 1; // 1-based
+      scores.set(id, (scores.get(id) ?? 0) + 1 / (k + rank));
+    }
+  }
+
+  return [...scores.entries()]
+    .map(([id, score]) => ({ id, score }))
+    .sort((a, b) => b.score - a.score);
+}

src/lib/search/search.ts

@@ -0,0 +1,77 @@
+// Cross-lecture search orchestrator (Phase 1). Embeds the query on-device,
+// pulls semantic candidates from the vector store, re-ranks them lexically, and
+// fuses the two signals with Reciprocal Rank Fusion. Returns segment-level hits
+// (with the matching signal tagged) that the UI can jump to. 100% on-device.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+import { getRepo } from '../db';
+import type { VectorHit } from '../db/repo';
+import { getEmbeddingEngine } from '../embedding';
+import { lexicalRank } from './lexical';
+import { rrf } from './rrf';
+import type { SearchHit, SearchOptions } from './types';
+
+/** How many semantic candidates to pull before re-ranking/fusing. */
+const CANDIDATE_LIMIT = 50;
+/** Default number of hits returned to the caller. */
+const DEFAULT_LIMIT = 30;
+
+/**
+ * Search the user's lectures for `query`, returning fused semantic + lexical
+ * hits at segment granularity (best first). Empty/whitespace queries return [].
+ */
+export async function searchLectures(
+  query: string,
+  opts?: SearchOptions,
+): Promise<SearchHit[]> {
+  const q = query.trim();
+  if (q.length === 0) return [];
+
+  // 1) Embed the query on-device (lazy-load the model on first use).
+  const engine = getEmbeddingEngine();
+  if (!engine.isLoaded()) await engine.loadModel();
+  const [vec] = await engine.embed([q], 'query');
+  if (!vec) return [];
+
+  // 2) Semantic candidates from the vector store (cosine, top CANDIDATE_LIMIT).
+  const cand = await getRepo().searchVectors(vec, {
+    courseId: opts?.courseId,
+    limit: CANDIDATE_LIMIT,
+  });
+  if (cand.length === 0) return [];
+
+  // O(1) candidate lookup by segmentId for the fusion map-back step.
+  const bySegment = new Map<string, VectorHit>();
+  for (const c of cand) bySegment.set(c.segmentId, c);
+
+  // 3) Two ranked id lists over the SAME candidate set.
+  //    - semantic: candidates are already cosine-desc from the repo.
+  const semanticIds = cand.map((c) => c.segmentId);
+  const semanticSet = new Set(semanticIds);
+  //    - lexical: keyword re-rank of just these candidates' texts.
+  const lexicalRanked = lexicalRank(
+    cand.map((c) => ({ id: c.segmentId, text: c.text })),
+    q,
+  );
+  const lexicalIds = lexicalRanked.map((r) => r.id);
+  const lexicalSet = new Set(lexicalIds);
+
+  // 4) Fuse the two orderings.
+  const fused = rrf([semanticIds, lexicalIds]);
+
+  // 5) Map fused ids back to their VectorHit and tag the matching signal.
+  const limit = opts?.limit ?? DEFAULT_LIMIT;
+  const hits: SearchHit[] = [];
+  for (const { id } of fused) {
+    const hit = bySegment.get(id);
+    if (!hit) continue;
+    const inSemantic = semanticSet.has(id);
+    const inLexical = lexicalSet.has(id);
+    const via: SearchHit['via'] = inSemantic && inLexical ? 'both' : inLexical ? 'lexical' : 'semantic';
+    hits.push({ ...hit, via });
+    if (hits.length >= limit) break;
+  }
+
+  return hits;
+}

src/lib/search/types.ts

@@ -0,0 +1,14 @@
+// Shared types for cross-lecture search (Phase 1).
+import type { VectorHit } from '../db/repo';
+
+/** A search result at segment granularity, with which signal(s) matched. */
+export interface SearchHit extends VectorHit {
+  via: 'semantic' | 'lexical' | 'both';
+}
+
+export interface SearchOptions {
+  /** Scope to a course (null = Unsorted); omit for all. */
+  courseId?: string | null;
+  /** Max hits to return. */
+  limit?: number;
+}

src/lib/transcription/engineImpl.native.ts

@@ -28,7 +28,10 @@
 
 import { initWhisper } from 'whisper.rn';
 import type { WhisperContext } from 'whisper.rn';
-import { Paths, File } from 'expo-file-system';
+import { Paths, File, Directory } from 'expo-file-system';
+// Progress-reporting downloads aren't in the new OO File API yet, so we use the
+// (still supported) legacy resumable downloader for the model fetch.
+import { createDownloadResumable } from 'expo-file-system/legacy';
 import type {
   ModelId,
   PcmAudio,
@@ -36,33 +39,95 @@ import type {
   TranscribeOptions,
 } from '../types';
 import { MODELS, recommendModel } from '../models/catalog';
+import { errorMessage } from '../errorMessage';
 import type { EngineCapabilities, TranscriptionEngine } from './engine';
 
 /** Loaded whisper.cpp contexts, keyed by model id. */
 const loaded = new Map<ModelId, WhisperContext>();
 
 /**
- * Resolve the on-device path of a model's ggml file.
+ * Where the official whisper.cpp ggml models are hosted. The catalog's `ggml`
+ * field is the exact filename in this repo (e.g. 'ggml-tiny.en.bin').
+ */
+const GGML_BASE_URL = 'https://huggingface.co/ggerganov/whisper.cpp/resolve/main';
+
+/**
+ * The on-device file handle for a model's ggml binary.
  *
  * Models live under `<documentDirectory>/models/<ggml-filename>`. The document
  * directory survives app restarts and OS storage pressure, which is what we want
  * for multi-hundred-MB model files.
- *
- * TODO: the model DOWNLOAD UI / fetcher is NOT built yet. This function only
- *       computes where the file *should* be; the .bin must already exist there
- *       (e.g. side-loaded during development) or loadModel() will reject. Wire a
- *       downloader (File.downloadFileAsync into Paths.document/'models') before
- *       shipping.
- *
- * whisper.rn's initWhisper expects a plain filesystem path. Expo's File `.uri`
- * is a `file://` URI, so we strip the scheme.
  */
-function modelPathFor(id: ModelId): string {
+function modelFile(id: ModelId): File {
-  const file = new File(Paths.document, 'models', MODELS[id].ggml);
+  return new File(Paths.document, 'models', MODELS[id].ggml);
-  // e.g. 'file:///data/.../Documents/models/ggml-tiny.en.bin' -> '/data/.../ggml-tiny.en.bin'
+}
+
+/** whisper.rn's initWhisper wants a plain path; Expo's `.uri` is `file://…`. */
+function plainPath(file: File): string {
   return file.uri.replace(/^file:\/\//, '');
 }
 
+/**
+ * Ensure a model's ggml file is present on disk, downloading it from Hugging
+ * Face if needed. Reports 0..1 download progress via `onProgress`. Downloads to
+ * a `.part` file and renames on success, so an interrupted download never leaves
+ * a truncated file that a later run would mistake for a complete model.
+ */
+async function ensureModelDownloaded(
+  id: ModelId,
+  onProgress?: (p: number) => void,
+): Promise<File> {
+  const dest = modelFile(id);
+  if (dest.exists) return dest;
+
+  // Make sure <documentDirectory>/models/ exists.
+  const dir = new Directory(Paths.document, 'models');
+  if (!dir.exists) dir.create({ intermediates: true, idempotent: true });
+
+  const url = `${GGML_BASE_URL}/${MODELS[id].ggml}`;
+  const partUri = `${dest.uri}.part`;
+
+  // Clean up any leftover .part from a previously-aborted attempt.
+  const part = new File(partUri);
+  if (part.exists) {
+    try {
+      part.delete();
+    } catch {
+      /* best effort */
+    }
+  }
+
+  const task = createDownloadResumable(url, partUri, {}, (p) => {
+    if (p.totalBytesExpectedToWrite > 0) {
+      onProgress?.(p.totalBytesWritten / p.totalBytesExpectedToWrite);
+    }
+  });
+
+  try {
+    const result = await task.downloadAsync();
+    if (!result || result.status !== 200) {
+      throw new Error(
+        `server returned ${result?.status ?? 'no response'}`,
+      );
+    }
+  } catch (e) {
+    try {
+      if (part.exists) part.delete();
+    } catch {
+      /* best effort */
+    }
+    throw new Error(
+      `Couldn't download the ${MODELS[id].label} model (~${MODELS[id].approxMB} MB): ` +
+        `${errorMessage(e)}. Check your connection and try again.`,
+    );
+  }
+
+  // Atomically put the completed file in place.
+  part.move(dest);
+  onProgress?.(1);
+  return dest;
+}
+
 export const engine: TranscriptionEngine = {
   platform: 'native',
 
@@ -81,13 +146,31 @@ export const engine: TranscriptionEngine = {
     };
   },
 
-  async loadModel(modelId: ModelId): Promise<void> {
+  async loadModel(
+    modelId: ModelId,
+    onProgress?: (p: number) => void,
+  ): Promise<void> {
     if (loaded.has(modelId)) return; // idempotent
 
-    const filePath = modelPathFor(modelId);
+    // Fetch the ggml file on first use (with download progress), then load it.
-    // initWhisper rejects if the file is missing/corrupt. See the TODO in
+    const file = await ensureModelDownloaded(modelId, onProgress);
-    // modelPathFor: until the download UI exists, the .bin must already be there.
+
-    const ctx = await initWhisper({ filePath });
+    let ctx: WhisperContext;
+    try {
+      ctx = await initWhisper({ filePath: plainPath(file) });
+    } catch (e) {
+      // The file is present but unusable (corrupt / truncated). Delete it so the
+      // next attempt re-downloads a clean copy rather than failing forever.
+      try {
+        if (file.exists) file.delete();
+      } catch {
+        /* best effort */
+      }
+      throw new Error(
+        `The ${MODELS[modelId].label} model couldn't be loaded and was cleared — ` +
+          `try again. (${errorMessage(e)})`,
+      );
+    }
     loaded.set(modelId, ctx);
   },
 
@@ -113,11 +196,15 @@ export const engine: TranscriptionEngine = {
     // Float32Array whose `.buffer` is exactly these samples.
     const pcm = audio.samples.slice();
 
+    // English-only models (".en") only do English — only pass language/translate
+    // for multilingual models (mirrors the web engine; keeps whisper.cpp from
+    // being asked to do something a .en model can't).
+    const multilingual = MODELS[opts.modelId].multilingual;
     // transcribeData returns { stop, promise }; await the promise.
     const { promise } = ctx.transcribeData(pcm.buffer, {
       // language undefined => whisper.rn auto-detects ('auto').
-      language: opts.language,
+      ...(multilingual && opts.language ? { language: opts.language } : {}),
-      translate: opts.translate ?? false,
+      ...(multilingual ? { translate: opts.translate ?? false } : {}),
     });
     const result = await promise;
 

src/lib/transcription/engineImpl.web.ts

@@ -40,7 +40,9 @@ interface AsrOutput {
 type AsrPipeline = (audio: Float32Array, opts: Record<string, unknown>) => Promise<AsrOutput>;
 interface PipelineOptions {
   device?: string;
-  dtype?: string;
+  // A single dtype for all sub-models, or a per-file map (encoder_model,
+  // decoder_model_merged, …) — Whisper's decoder needs special handling on WASM.
+  dtype?: string | Record<string, string>;
   progress_callback?: (e: { status?: string; progress?: number }) => void;
 }
 interface TransformersModule {
@@ -64,8 +66,21 @@ async function lib(): Promise<TransformersModule> {
 const loaded = new Map<ModelId, AsrPipeline>();
 let cachedBackend: Backend | undefined;
 
+/**
+ * Mobile browsers expose WebGPU but their drivers are flaky for sustained ML
+ * inference — fp16 Whisper tends to crash the GPU process after a chunk or two.
+ * On mobile we deliberately use (cross-origin-isolated, multi-threaded) WASM,
+ * which is slower but reliable. Desktop keeps WebGPU.
+ */
+function isMobileWeb(): boolean {
+  if (typeof navigator === 'undefined') return false;
+  const ua = navigator.userAgent || '';
+  return /Android|iPhone|iPad|iPod|Mobile|Silk|Kindle/i.test(ua);
+}
+
 async function detectWebGpu(): Promise<boolean> {
   try {
+    if (isMobileWeb()) return false;
     if (typeof navigator === 'undefined' || !('gpu' in navigator)) return false;
     const gpu = (navigator as { gpu?: { requestAdapter(): Promise<unknown> } }).gpu;
     const adapter = await gpu?.requestAdapter();
@@ -98,10 +113,17 @@ export const engine: TranscriptionEngine = {
     const { pipeline } = await lib();
     const webgpu = (await resolveBackend()) === 'webgpu';
     const asr = await pipeline('automatic-speech-recognition', MODELS[modelId].webRepo, {
-      // WebGPU + fp16 when available; otherwise 8-bit weights on WASM, which
-      // stays small to download and runs acceptably on a plain CPU.
       device: webgpu ? 'webgpu' : 'wasm',
-      dtype: webgpu ? 'fp16' : 'q8',
+      // WebGPU: fp16 (fast, small). WASM: an explicit fp32 decoder. The default
+      // quantized decoders (q8/q4) use `MatMulNBits` ops that onnxruntime-web
+      // (transformers.js 4.2.0) FAILS to load on the WASM backend with
+      // "Missing required scale" — verified across Xenova + onnx-community
+      // repos. fp32 avoids those ops and actually runs on any CPU (the WASM
+      // path is what mobile and no-WebGPU devices use). Larger download, but it
+      // works; q8 simply does not load on WASM here.
+      dtype: webgpu
+        ? 'fp16'
+        : { encoder_model: 'fp32', decoder_model_merged: 'fp32' },
       progress_callback: (e) => {
         if (e.status === 'progress' && e.progress != null) onProgress?.(e.progress / 100);
       },
@@ -117,21 +139,43 @@ export const engine: TranscriptionEngine = {
     const asr = loaded.get(opts.modelId);
     if (!asr) throw new Error(`Model "${opts.modelId}" is not loaded; call loadModel() first.`);
 
-    const out = await asr(audio.samples, {
+    // English-only Whisper models (".en") reject `language`/`task` — transformers.js
+    // throws "Cannot specify `task` or `language` for an English-only model" if
+    // either is passed. Only multilingual models accept them (translation also
+    // requires a multilingual model). So we add those keys ONLY when multilingual,
+    // and never pass an explicit `undefined` language (which trips the same check).
+    const genOpts: Record<string, unknown> = {
       return_timestamps: true,
       // One window at a time; 30s matches Whisper's frame so it won't re-chunk.
       chunk_length_s: 30,
-      language: opts.language,
+    };
-      task: opts.translate ? 'translate' : 'transcribe',
+    if (MODELS[opts.modelId].multilingual) {
-    });
+      genOpts.task = opts.translate ? 'translate' : 'transcribe';
+      if (opts.language) genOpts.language = opts.language;
+    }
 
+    const out = await asr(audio.samples, genOpts);
+
+    // Whisper sometimes omits the END timestamp of the last utterance in a
+    // window (transformers.js yields timestamp[1] === null). Don't collapse that
+    // to a zero-length [t, t] segment (it breaks stitch seams + citation spans):
+    // estimate the end from the next chunk's start, else the window's duration.
+    const chunks = out.chunks ?? [];
+    const chunkDurationSec = audio.samples.length / audio.sampleRate;
     const segments: Segment[] = [];
-    for (const c of out.chunks ?? []) {
+    for (let i = 0; i < chunks.length; i++) {
-      const [start, end] = c.timestamp;
+      const c = chunks[i]!;
+      const [start, rawEnd] = c.timestamp;
       if (start == null) continue;
       const text = c.text.trim();
       if (text.length === 0) continue;
-      segments.push({ start, end: end ?? start, text });
+      let end = rawEnd;
+      if (end == null) {
+        const nextStart = chunks[i + 1]?.timestamp[0];
+        end = nextStart != null && nextStart > start ? nextStart : chunkDurationSec;
+      }
+      if (end < start) end = chunkDurationSec;
+      segments.push({ start, end, text });
     }
     return segments;
   },

src/lib/transcription/pipeline.test.ts

@@ -127,16 +127,21 @@ describe('transcribe pipeline', () => {
     const transcribing = events.filter((e) => e.stage === 'transcribing');
     expect(transcribing.length).toBeGreaterThan(0);
 
-    // Monotonically non-decreasing, strictly bounded in (0, 1].
+    // The first transcribing event is a kickoff fired BEFORE the first chunk so
-    for (let i = 0; i < transcribing.length; i++) {
+    // the UI leaves the "loading" stage immediately: progress 0, no partial yet.
-      const e = transcribing[i]!;
+    expect(transcribing[0]!.progress).toBe(0);
-      expect(e.progress).toBeGreaterThan(0);
+    expect(transcribing[0]!.partial).toEqual([]);
+
+    // Every event carries the total chunk count; progress stays within [0, 1].
+    for (const e of transcribing) {
+      expect(e.chunkCount).toBeGreaterThan(0);
+      expect(e.progress).toBeGreaterThanOrEqual(0);
       expect(e.progress).toBeLessThanOrEqual(1);
-      if (i > 0) {
-        expect(e.progress).toBeGreaterThan(transcribing[i - 1]!.progress);
-      }
     }
-    // Final transcribing event hits exactly 1.
+    // After the kickoff, per-chunk progress strictly increases and ends at 1.
+    for (let i = 2; i < transcribing.length; i++) {
+      expect(transcribing[i]!.progress).toBeGreaterThan(transcribing[i - 1]!.progress);
+    }
     expect(transcribing[transcribing.length - 1]!.progress).toBe(1);
 
     // The growing partial should reach the final segment count on the last tick.

src/lib/transcription/pipeline.ts

@@ -26,6 +26,10 @@ export interface TranscribeProgress {
    * grows after each chunk during 'transcribing'.
    */
   partial: Segment[];
+  /** Index of the chunk currently being / just transcribed (0-based). */
+  chunkIndex?: number;
+  /** Total number of chunks planned for this audio. */
+  chunkCount?: number;
 }
 
 export interface TranscribeParams {
@@ -80,6 +84,18 @@ export async function transcribe(params: TranscribeParams): Promise<Segment[]> {
   // 3. Transcribe each window. `perChunk[i]` holds chunk-local segments for
   //    plan[i]; the stitcher offsets and merges them.
   const perChunk: Segment[][] = [];
+
+  // Flip to the 'transcribing' stage immediately, BEFORE the first (often slow)
+  // chunk runs — otherwise the UI sits at "Loading model… 100%" for the whole
+  // first inference and looks frozen/crashed. progress 0/N, no partial yet.
+  onProgress?.({
+    stage: 'transcribing',
+    progress: 0,
+    partial: [],
+    chunkIndex: 0,
+    chunkCount: plan.length,
+  });
+
   for (let i = 0; i < plan.length; i++) {
     // Cooperative cancellation: check before each (potentially long) chunk.
     if (signal?.aborted) throw new DOMException('Aborted', 'AbortError');
@@ -92,6 +108,13 @@ export async function transcribe(params: TranscribeParams): Promise<Segment[]> {
     };
 
     const segs = await engine.transcribeChunk(chunkAudio, options);
+
+    // Re-check AFTER the (long) inference: the top-of-loop check can't catch a
+    // cancel that happens DURING transcribeChunk — and for single-chunk audio
+    // the loop body runs once, so without this a late Cancel would still fall
+    // through and save the transcript.
+    if (signal?.aborted) throw new DOMException('Aborted', 'AbortError');
+
     perChunk.push(segs);
 
     // Emit a growing absolute-time partial: stitch everything decoded so far.
@@ -99,6 +122,8 @@ export async function transcribe(params: TranscribeParams): Promise<Segment[]> {
       stage: 'transcribing',
       progress: (i + 1) / plan.length,
       partial: stitchSegments(perChunk, plan.slice(0, i + 1), { overlapSec }),
+      chunkIndex: i + 1,
+      chunkCount: plan.length,
     });
   }
 

src/lib/types.ts

@@ -12,6 +12,12 @@ export const WHISPER_SAMPLE_RATE = 16000 as const;
  * pipeline offsets and stitches them into absolute time.
  */
 export interface Segment {
+  /**
+   * Stable per-segment id, assigned by the storage layer on create()/update().
+   * Engine output omits it; it is the durable anchor for citations, vectors,
+   * flashcards and calendar links (see ROADMAP Phase 0/1).
+   */
+  id?: string;
   /** Inclusive start time in seconds. */
   start: number;
   /** Exclusive end time in seconds. */

src/stores/aiStore.ts

@@ -0,0 +1,159 @@
+// Session/UI state for the optional generative ("ask your lectures") layer
+// (Phase 5). Holds the BYO-key cloud config (if any), the on-device model load
+// status, and which backend is selected, and exposes a single ask() that wires
+// retrieval -> generation via askLectures.
+//
+// The feature is GATED BY AVAILABILITY, not a manual switch: `enabled` defaults
+// true, but ask() still degrades to the search-only fallback when no engine is
+// available (no WebGPU and no cloud key). Raw audio/transcripts never leave the
+// device — only the question + retrieved snippets do (cloud path).
+//
+// Persistence: on web the cloud config is saved to localStorage under
+// 'wisp:ai'; native keeps it in-memory (no localStorage). Everything here is
+// defensive — refreshAvailability never throws.
+import { create } from 'zustand';
+
+import {
+  getGenerationEngine,
+  type CloudConfig,
+  type GenerationEngine,
+  type RagAnswer,
+} from '@/lib/generation';
+import { askLectures } from '@/lib/generation/rag';
+
+const STORAGE_KEY = 'wisp:ai';
+
+type ModelStatus = 'idle' | 'loading' | 'ready';
+type EngineKind = 'webllm' | 'cloud' | 'none';
+
+interface PersistedState {
+  enabled: boolean;
+  cloud?: CloudConfig;
+}
+
+/** Web-only localStorage handle (undefined on native / SSR). */
+function ls(): Storage | undefined {
+  try {
+    return typeof localStorage !== 'undefined' ? localStorage : undefined;
+  } catch {
+    return undefined;
+  }
+}
+
+/** Read persisted enabled + cloud config from localStorage (web only). */
+function loadPersisted(): PersistedState {
+  const store = ls();
+  if (!store) return { enabled: true };
+  try {
+    const raw = store.getItem(STORAGE_KEY);
+    if (!raw) return { enabled: true };
+    const parsed = JSON.parse(raw) as Partial<PersistedState>;
+    return {
+      enabled: parsed.enabled ?? true,
+      cloud: parsed.cloud,
+    };
+  } catch {
+    return { enabled: true };
+  }
+}
+
+/** Persist (or clear) the cloud config + enabled flag on web. No-op on native. */
+function persist(state: PersistedState): void {
+  const store = ls();
+  if (!store) return;
+  try {
+    store.setItem(STORAGE_KEY, JSON.stringify(state));
+  } catch {
+    /* quota / private mode — ignore, just stay in-memory */
+  }
+}
+
+const initial = loadPersisted();
+
+interface AiState {
+  /** Feature flag — defaults true; real gating is via availability. */
+  enabled: boolean;
+  /** BYO-key cloud config, if the user supplied one. */
+  cloud?: CloudConfig;
+  /** On-device model load status (WebLLM). */
+  modelStatus: ModelStatus;
+  /** WebLLM download progress in [0,1]. */
+  progress: number;
+  /** Which backend the current config resolves to. */
+  engineKind: EngineKind;
+
+  /** Set (or clear, with no arg) the cloud config; persists on web. */
+  setCloud: (cfg?: CloudConfig) => void;
+  /** Recompute engineKind and probe availability. Never throws. */
+  refreshAvailability: () => Promise<void>;
+  /** Resolve an engine, loading the on-device model if needed. */
+  ensureReady: () => Promise<GenerationEngine>;
+  /** Retrieve + (optionally) generate a grounded, cited answer. */
+  ask: (
+    question: string,
+    opts?: { courseId?: string | null; signal?: AbortSignal },
+  ) => Promise<RagAnswer>;
+}
+
+export const useAi = create<AiState>((set, get) => ({
+  enabled: initial.enabled,
+  cloud: initial.cloud,
+  modelStatus: 'idle',
+  progress: 0,
+  // Resolve the initial backend synchronously from the persisted config.
+  engineKind: getGenerationEngine(initial.cloud).kind,
+
+  setCloud: (cfg) => {
+    persist({ enabled: get().enabled, cloud: cfg });
+    // A cloud config swap invalidates any in-progress on-device load state.
+    set({
+      cloud: cfg,
+      engineKind: getGenerationEngine(cfg).kind,
+      modelStatus: 'idle',
+      progress: 0,
+    });
+  },
+
+  refreshAvailability: async () => {
+    try {
+      const engine = getGenerationEngine(get().cloud);
+      set({ engineKind: engine.kind });
+      // Probe availability (WebGPU adapter / configured key). The result isn't
+      // stored beyond engineKind; ask()/ensureReady() re-check at call time.
+      await engine.isAvailable();
+    } catch {
+      // Stay defensive: never throw out of availability refresh.
+      set({ engineKind: 'none' });
+    }
+  },
+
+  ensureReady: async () => {
+    const engine = getGenerationEngine(get().cloud);
+    set({ engineKind: engine.kind });
+    // Only the on-device WebLLM engine has a model to download — and only when
+    // it can actually run here (WebGPU). On a machine without WebGPU we skip the
+    // (doomed, ~1GB) load and let askLectures degrade to search-only.
+    if (
+      engine.kind === 'webllm' &&
+      !engine.isLoaded() &&
+      (await engine.isAvailable())
+    ) {
+      set({ modelStatus: 'loading', progress: 0 });
+      try {
+        await engine.loadModel((p) => set({ progress: p }));
+        set({ modelStatus: 'ready', progress: 1 });
+      } catch (err) {
+        set({ modelStatus: 'idle', progress: 0 });
+        throw err;
+      }
+    } else {
+      set({ modelStatus: 'ready' });
+    }
+    return engine;
+  },
+
+  ask: async (question, opts) => {
+    const engine = await get().ensureReady();
+    return askLectures(question, engine, opts);
+  },
+}));

src/stores/coursesStore.ts

@@ -0,0 +1,43 @@
+// Courses list state, backed by the local StorageRepo. UI/session state only.
+import { create } from 'zustand';
+
+import { getRepo, type Course, type CourseDraft } from '@/lib/db';
+
+interface CoursesState {
+  items: Course[];
+  loading: boolean;
+  refresh: () => Promise<void>;
+  createCourse: (draft: CourseDraft) => Promise<Course>;
+  rename: (id: string, name: string) => Promise<void>;
+  remove: (id: string) => Promise<void>;
+}
+
+export const useCourses = create<CoursesState>((set, get) => ({
+  items: [],
+  loading: false,
+
+  refresh: async () => {
+    set({ loading: true });
+    try {
+      set({ items: await getRepo().listCourses() });
+    } finally {
+      set({ loading: false });
+    }
+  },
+
+  createCourse: async (draft) => {
+    const course = await getRepo().createCourse(draft);
+    await get().refresh();
+    return course;
+  },
+
+  rename: async (id, name) => {
+    await getRepo().updateCourse(id, { name });
+    await get().refresh();
+  },
+
+  remove: async (id) => {
+    await getRepo().deleteCourse(id);
+    await get().refresh();
+  },
+}));

src/stores/embeddingStore.ts

@@ -0,0 +1,129 @@
+// Drives on-device indexing for semantic search (Phase 1): loads the embedding
+// model, embeds each transcript's segments into unit vectors, and persists them
+// to the StorageRepo's vector store. UI/session state only — the repo is the
+// record of truth for what's been embedded.
+import { create } from 'zustand';
+
+import { getRepo } from '@/lib/db';
+import { EMBED_MODEL, getEmbeddingEngine } from '@/lib/embedding';
+
+type Status = 'idle' | 'indexing' | 'ready';
+
+interface EmbeddingState {
+  status: Status;
+  /** Build progress in [0,1] (0.2 = model load, 0.8 = embedding). */
+  progress: number;
+  /** Count of transcripts with no vectors for the active model. */
+  pending: number;
+
+  /** Refresh `pending` from the repo (cheap; no model load). */
+  refreshPending: () => Promise<void>;
+  /** Embed every un-embedded transcript. Loads the model first. */
+  buildIndex: () => Promise<void>;
+  /** Embed a single transcript (used right after a new transcription). */
+  embedOne: (transcriptId: string) => Promise<void>;
+}
+
+/**
+ * Embed and persist vectors for one transcript. Replaces any existing vectors
+ * for it (upsertVectors semantics). No-op if the transcript or its segments are
+ * gone. Caller is responsible for model loading and status/progress bookkeeping.
+ */
+async function embedTranscript(transcriptId: string): Promise<void> {
+  const repo = getRepo();
+  const t = await repo.get(transcriptId);
+  if (!t) return;
+  if (t.segments.length === 0) {
+    // Nothing to embed, but record an (empty) vector set so it stops counting
+    // as "unembedded" for this model.
+    await repo.upsertVectors(transcriptId, EMBED_MODEL, []);
+    return;
+  }
+  const eng = getEmbeddingEngine();
+  // Embed in small batches, yielding to the event loop between them. The WASM
+  // backend runs synchronously on the MAIN thread, so embedding a whole lecture
+  // (100s of segments) in one call froze the UI for seconds — which, right after
+  // a transcription, looked like the transcript screen was "stuck loading" and
+  // starved any follow-up transcription. Per-text vectors are identical
+  // regardless of batch size (e5 pools each text independently).
+  const texts = t.segments.map((s) => s.text);
+  const BATCH = 16;
+  const vecs: Float32Array[] = [];
+  for (let i = 0; i < texts.length; i += BATCH) {
+    const out = await eng.embed(texts.slice(i, i + BATCH), 'passage');
+    for (const v of out) vecs.push(v);
+    // Hand the thread back so navigation, rendering and other work can run.
+    await new Promise<void>((resolve) => setTimeout(resolve, 0));
+  }
+  await repo.upsertVectors(
+    transcriptId,
+    EMBED_MODEL,
+    t.segments.map((s, j) => ({
+      segmentId: s.id ?? String(j),
+      start: s.start,
+      end: s.end,
+      text: s.text,
+      vector: vecs[j] ?? new Float32Array(eng.dim),
+    })),
+  );
+}
+
+export const useEmbedding = create<EmbeddingState>((set, get) => ({
+  status: 'idle',
+  progress: 0,
+  pending: 0,
+
+  refreshPending: async () => {
+    const ids = await getRepo().unembeddedIds(EMBED_MODEL);
+    set({ pending: ids.length });
+  },
+
+  buildIndex: async () => {
+    // Guard against concurrent runs.
+    if (get().status === 'indexing') return;
+    set({ status: 'indexing', progress: 0 });
+    try {
+      const eng = getEmbeddingEngine();
+      // Model load is the first 20% of the bar.
+      await eng.loadModel((p) => set({ progress: p * 0.2 }));
+
+      const ids = await getRepo().unembeddedIds(EMBED_MODEL);
+      if (ids.length === 0) {
+        set({ status: 'ready', progress: 1, pending: 0 });
+        return;
+      }
+
+      for (let i = 0; i < ids.length; i++) {
+        const id = ids[i];
+        if (id === undefined) continue;
+        await embedTranscript(id);
+        // Remaining 80% spread across the transcripts.
+        set({ progress: 0.2 + 0.8 * ((i + 1) / ids.length) });
+      }
+
+      set({ status: 'ready', progress: 1, pending: 0 });
+    } catch (err) {
+      // Surface progress reset but don't crash callers; leave pending intact so
+      // a retry can pick up where it left off.
+      set({ status: 'idle', progress: 0 });
+      throw err;
+    }
+  },
+
+  embedOne: async (transcriptId) => {
+    // Guard against clobbering an in-flight full build.
+    if (get().status === 'indexing') return;
+    set({ status: 'indexing' });
+    try {
+      const eng = getEmbeddingEngine();
+      // Lazy model load (no progress weighting for a single transcript).
+      if (!eng.isLoaded()) await eng.loadModel();
+      await embedTranscript(transcriptId);
+      const pending = (await getRepo().unembeddedIds(EMBED_MODEL)).length;
+      set({ status: 'ready', pending });
+    } catch (err) {
+      set({ status: 'idle' });
+      throw err;
+    }
+  },
+}));

src/stores/transcribeStore.ts

@@ -6,10 +6,12 @@ import { create } from 'zustand';
 
 import { getDecoder, type AudioFileInput } from '@/lib/audio';
 import { getRepo } from '@/lib/db';
+import { errorMessage } from '@/lib/errorMessage';
 import { DEFAULT_MODEL } from '@/lib/models/catalog';
 import { getEngine } from '@/lib/transcription';
 import { transcribe } from '@/lib/transcription/pipeline';
 import type { ModelId, Segment } from '@/lib/types';
+import { useEmbedding } from './embeddingStore';
 
 type Status = 'idle' | 'loading' | 'transcribing' | 'done' | 'error';
 
@@ -17,6 +19,8 @@ interface StartOptions {
   title?: string;
   language?: string;
   translate?: boolean;
+  courseId?: string | null;
+  lectureDate?: number;
 }
 
 interface TranscribeState {
@@ -24,6 +28,9 @@ interface TranscribeState {
   stage?: 'loading' | 'transcribing';
   progress: number;
   partial: Segment[];
+  /** During 'transcribing': how many chunks are done and how many total. */
+  chunkIndex?: number;
+  chunkCount?: number;
   error?: string;
   modelId: ModelId;
   lastTranscriptId?: string;
@@ -63,6 +70,8 @@ export const useTranscribe = create<TranscribeState>((set, get) => ({
       stage: 'loading',
       progress: 0,
       partial: [],
+      chunkIndex: undefined,
+      chunkCount: undefined,
       error: undefined,
       lastTranscriptId: undefined,
       audioUrl: makeAudioUrl(input),
@@ -78,7 +87,14 @@ export const useTranscribe = create<TranscribeState>((set, get) => ({
         engine: getEngine(),
         signal: abort.signal,
         onProgress: (p) =>
-          set({ stage: p.stage, progress: p.progress, partial: p.partial, status: p.stage }),
+          set({
+            stage: p.stage,
+            progress: p.progress,
+            partial: p.partial,
+            status: p.stage,
+            chunkIndex: p.chunkIndex,
+            chunkCount: p.chunkCount,
+          }),
       });
 
       const durationSec = pcm.samples.length / pcm.sampleRate;
@@ -88,8 +104,27 @@ export const useTranscribe = create<TranscribeState>((set, get) => ({
         modelId,
         language: opts?.language,
         segments,
+        courseId: opts?.courseId ?? null,
+        ...(opts?.lectureDate !== undefined ? { lectureDate: opts.lectureDate } : {}),
       });
 
+      // Persist the source audio so it can be replayed after a reload. Best-effort:
+      // a quota/unsupported failure must not lose the transcript itself.
+      try {
+        await getRepo().putMedia(saved.id, {
+          data: input.data,
+          uri: input.uri,
+          mime: guessMime(input.name),
+        });
+      } catch (e) {
+        console.warn('[wisp] could not persist source audio:', e);
+      }
+
+      // Index the new transcript for semantic search in the background. Lazy
+      // model load happens inside embedOne; this must never block or fail the
+      // transcription itself, so it's fire-and-forget with a swallowed error.
+      void useEmbedding.getState().embedOne(saved.id).catch(() => {});
+
       set({ status: 'done', progress: 1, partial: segments, lastTranscriptId: saved.id });
       return saved.id;
     } catch (err) {
@@ -97,7 +132,7 @@ export const useTranscribe = create<TranscribeState>((set, get) => ({
         set({ status: 'idle', progress: 0 });
         return undefined;
       }
-      set({ status: 'error', error: err instanceof Error ? err.message : String(err) });
+      set({ status: 'error', error: errorMessage(err) });
       return undefined;
     }
   },
@@ -118,3 +153,14 @@ function defaultTitle(): string {
   const pad = (n: number) => String(n).padStart(2, '0');
   return `Transcript ${d.getFullYear()}-${pad(d.getMonth() + 1)}-${pad(d.getDate())} ${pad(d.getHours())}:${pad(d.getMinutes())}`;
 }
+
+/** Best-effort MIME from a file name extension (for persisted media). */
+function guessMime(name?: string): string {
+  const ext = (name?.split('.').pop() || '').toLowerCase();
+  const map: Record<string, string> = {
+    mp3: 'audio/mpeg', m4a: 'audio/mp4', aac: 'audio/aac', wav: 'audio/wav',
+    ogg: 'audio/ogg', oga: 'audio/ogg', flac: 'audio/flac', webm: 'audio/webm',
+    mp4: 'video/mp4', mov: 'video/quicktime', mkv: 'video/x-matroska',
+  };
+  return map[ext] || 'application/octet-stream';
+}

src/stores/transcriptsStore.ts

@@ -1,34 +1,58 @@
-// Library state: the saved transcripts list + search, backed by the local
+// Library state: the saved transcripts list + search + course filter, backed by
-// StorageRepo (Dexie on web, expo-sqlite on native). UI/session state only —
+// the local StorageRepo. UI/session state only — the repo is the record of truth.
-// the record of truth lives in the repo.
 import { create } from 'zustand';
 
 import { getRepo, type TranscriptMeta } from '@/lib/db';
 
+/** 'all' = every transcript · null = Unsorted · string = a specific course id. */
+export type CourseFilter = 'all' | null | string;
+
+// Monotonic token so overlapping refreshes (focus + typing + filter toggles all
+// call refresh()) can't resolve out of order and clobber the list with a stale
+// query/filter's results. Only the latest-issued refresh is allowed to commit.
+let refreshToken = 0;
+
 interface TranscriptsState {
   items: TranscriptMeta[];
   loading: boolean;
   query: string;
+  courseFilter: CourseFilter;
   refresh: () => Promise<void>;
   setQuery: (q: string) => Promise<void>;
+  setCourseFilter: (f: CourseFilter) => Promise<void>;
   remove: (id: string) => Promise<void>;
   rename: (id: string, title: string) => Promise<void>;
+  reassign: (id: string, courseId: string | null) => Promise<void>;
 }
 
 export const useTranscripts = create<TranscriptsState>((set, get) => ({
   items: [],
   loading: false,
   query: '',
+  courseFilter: 'all',
 
   refresh: async () => {
+    const my = ++refreshToken;
     set({ loading: true });
     try {
       const repo = getRepo();
       const q = get().query.trim();
-      const items = q ? await repo.search(q) : await repo.list();
+      const cf = get().courseFilter;
+      let items: TranscriptMeta[];
+      if (q) {
+        // Search across everything, then narrow to the active course filter.
+        items = await repo.search(q);
+        if (cf !== 'all') items = items.filter((t) => (t.courseId ?? null) === cf);
+      } else {
+        items = cf === 'all' ? await repo.list() : await repo.listByCourse(cf);
+      }
+      // A newer refresh was issued while we awaited — discard our stale result.
+      if (my !== refreshToken) return;
       set({ items });
     } finally {
-      set({ loading: false });
+      // Only the latest refresh clears the spinner, so it doesn't flicker off
+      // while a newer in-flight refresh is still running.
+      if (my === refreshToken) set({ loading: false });
     }
   },
 
@@ -37,6 +61,11 @@ export const useTranscripts = create<TranscriptsState>((set, get) => ({
     await get().refresh();
   },
 
+  setCourseFilter: async (f) => {
+    set({ courseFilter: f });
+    await get().refresh();
+  },
+
   remove: async (id) => {
     await getRepo().remove(id);
     await get().refresh();
@@ -46,4 +75,9 @@ export const useTranscripts = create<TranscriptsState>((set, get) => ({
     await getRepo().update(id, { title });
     await get().refresh();
   },
+
+  reassign: async (id, courseId) => {
+    await getRepo().reassign(id, courseId);
+    await get().refresh();
+  },
 }));