Bug-hunt sweep: fix 13 verified issues (correctness + perf)

From a multi-agent bug-hunt + adversarial verification pass (0 critical/
high; 15 mediums). Fixed 13; 2 deferred as bigger refactors.

Correctness:
- srs: "Hard" no longer overshoots "Good" for reviewed cards (reps>=2). It
  compounded ease AND x1.2; now grows x1.2 off the previous interval only.
  + regression test.
- enrich/dates: stop reading the modal verb "may" as the month May (phantom
  calendar events). "may" needs an ordinal/year/date-preposition now.
- enrich/bib: disambiguate colliding BibTeX keys (smith2020, smith2020a, …)
  — duplicates corrupted reference-manager imports.
- learn/glossary: junk-term guard used && (dead); now || so all-stopword
  terms like "there" are actually skipped.
- transcription/engineImpl.web: don't collapse Whisper's null end-timestamp
  to a zero-length [t,t] segment; estimate from the next chunk or window
  duration (fixes citation/seam anchors).
- transcription/pipeline: re-check the abort signal AFTER each chunk so
  Cancel works on single-chunk audio (it previously still saved).
- db/repo.native: use withExclusiveTransactionAsync for reassign /
  deleteCourse / upsertVectors / createFlashcards (withTransactionAsync is
  not isolated on a shared connection → interleaved/half-applied writes).
- db/repo.native: don't cache a rejected open/migrate promise — a transient
  first-open failure no longer bricks storage for the whole session.
- stores/transcriptsStore: sequence-guard refresh() so overlapping
  focus/typing/filter refreshes can't resolve out of order and show stale
  results.
- audio/wav: decode WAVE_FORMAT_EXTENSIBLE (0xFFFE) via its SubFormat GUID
  + add 24-bit PCM — common ffmpeg/Windows WAVs no longer hard-fail native
  import. + tests.

Performance / footprint:
- audio/decode.native: decode straight to 16kHz (decodeAudioData sampleRate
  hint) so the JS side never holds a full-rate buffer or runs the resample
  loop — big memory/OOM win on long lectures.
- models/catalog display: Settings model sizes are now backend-aware (the
  no-GPU WASM path pulls ~2x fp32 weights; was advertising ~half).

Feature gap:
- download: native exports were silent no-ops. New download.native.ts writes
  to cache + opens the share sheet (expo-sharing); transcript/ICS/Anki-CSV/
  BibTeX/RIS exports now work on device.

Deferred (bigger): lexical-search recall ceiling (needs a full-corpus rank
path on both repos); triple in-memory copy of the encoded file during
transcribe (media is keyed by a not-yet-existing transcript id).

Validated: tsc clean, 282 tests pass, web export clean (native deps not
bundled), arm64 APK compiles with expo-sharing.

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

bun.lock

@@ -18,6 +18,7 @@
         "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",
@@ -943,6 +944,8 @@
 
     "expo-server": ["expo-server@56.0.5", "", {}, "sha512-SmM2p2g3Jrktpiazcst+OxhjSzOHXKAY4BPURHYHXvApzzoybMmrNF4IEZ8DKZ145BhSe4ydAmlEFCRTsdtgUQ=="],
 
+    "expo-sharing": ["expo-sharing@56.0.17", "", { "dependencies": { "@expo/config-plugins": "^56.0.8", "@expo/config-types": "^56.0.5", "@expo/plist": "^0.7.0" }, "peerDependencies": { "expo": "*", "react": "*", "react-native": "*" } }, "sha512-FqN7/UBau0PJ2O8OeMYS/fwE+6UMtdoDeGxsRoaileK0w30bXC92MuT7z4ujnk4mF9ZZBjS8axbGOgrZ6JWBEA=="],
+
     "expo-splash-screen": ["expo-splash-screen@56.0.10", "", { "dependencies": { "@expo/config-plugins": "~56.0.8", "@expo/image-utils": "^0.10.1", "xml2js": "0.6.0" }, "peerDependencies": { "expo": "*" } }, "sha512-vDIlo8hzt9HlCZQ0kSY66v83D1WEXOJbVMeyPDfXDu9tbDdPMNUyDpi4WGJXikAjxnAKfbt5Mv5NnEbxINy+VA=="],
 
     "expo-sqlite": ["expo-sqlite@56.0.5", "", { "dependencies": { "await-lock": "^2.2.2" }, "peerDependencies": { "expo": "*", "react": "*", "react-native": "*" } }, "sha512-wHYRVLS5nUFEtli45wHaO+RjlRY8sQXyOSgENVk6I4zq7+FgySqjOk3YOYW6IKIMwhj5XzjMJO+pY8xKUy73Kw=="],
@@ -1763,7 +1766,7 @@
 
     "xml2js": ["xml2js@0.6.0", "", { "dependencies": { "sax": ">=0.6.0", "xmlbuilder": "~11.0.0" } }, "sha512-eLTh0kA8uHceqesPqSE+VvO1CDDJWMwlQfB6LuN6T8w6MaDJ8Txm8P7s5cHD0miF0V+GGTZrDQfxPZQVsur33w=="],
 
-    "xmlbuilder": ["xmlbuilder@11.0.1", "", {}, "sha512-fDlsI/kFEx7gLvbecc0/ohLG50fugQp8ryHzMTuW9vSa1GJ0XYWKnhsUx7oie3G98+r56aTQIUB4kht42R3JvA=="],
+    "xmlbuilder": ["xmlbuilder@15.1.1", "", {}, "sha512-yMqGBqtXyeN1e3TGYvgNgDVZ3j84W4cwkOXQswghol6APgZWaff9lnbvN7MHYJOiXsvGPXtjTYJEiC9J2wv9Eg=="],
 
     "y18n": ["y18n@5.0.8", "", {}, "sha512-0pfFzegeDWJHJIAmTLRP2DwHjdF5s7jo9tuztdQxAhINCdvS+3nGINqPd00AphqJR/0LhANUS6/+7SCb98YOfA=="],
 
@@ -1809,8 +1812,6 @@
 
     "@expo/metro-config/hermes-parser": ["hermes-parser@0.33.3", "", { "dependencies": { "hermes-estree": "0.33.3" } }, "sha512-Yg3HgaG4CqgyowtYjX/FsnPAuZdHOqSMtnbpylbptsQ9nwwSKsy6uRWcGO5RK0EqiX12q8HvDWKgeAVajRO5DA=="],
 
-    "@expo/plist/xmlbuilder": ["xmlbuilder@15.1.1", "", {}, "sha512-yMqGBqtXyeN1e3TGYvgNgDVZ3j84W4cwkOXQswghol6APgZWaff9lnbvN7MHYJOiXsvGPXtjTYJEiC9J2wv9Eg=="],
-
     "@expo/ws-tunnel/ws": ["ws@8.21.0", "", { "peerDependencies": { "bufferutil": "^4.0.1", "utf-8-validate": ">=5.0.2" }, "optionalPeers": ["bufferutil", "utf-8-validate"] }, "sha512-Vsp28b7DRcimFQvrqu2Wek3z1iYxDCWqHYB8Qsnk/S4RfaCQzPGPyBNuVjJV3cd6UiKtUtp6sNM77gWvzcCH+g=="],
 
     "@react-native/babel-preset/@react-native/babel-plugin-codegen": ["@react-native/babel-plugin-codegen@0.86.0", "", { "dependencies": { "@babel/traverse": "^7.29.0", "@react-native/codegen": "0.86.0" } }, "sha512-qdsABWNW7uTll90l4Vh03gjeyu3WVDi2CyiiyvYGMRDcoYbjbQi6df3BMAm9lQI2yslZ1T14LlDDAsgTwNxplA=="],
@@ -1917,8 +1918,6 @@
 
     "plist/@xmldom/xmldom": ["@xmldom/xmldom@0.9.10", "", {}, "sha512-A9gOqLdi6cV4ibazAjcQufGj0B1y/vDqYrcuP6d/6x8P27gRS8643Dj9o1dEKtB6O7fwxb2FgBmJS2mX7gpvdw=="],
 
-    "plist/xmlbuilder": ["xmlbuilder@15.1.1", "", {}, "sha512-yMqGBqtXyeN1e3TGYvgNgDVZ3j84W4cwkOXQswghol6APgZWaff9lnbvN7MHYJOiXsvGPXtjTYJEiC9J2wv9Eg=="],
-
     "pretty-format/ansi-styles": ["ansi-styles@5.2.0", "", {}, "sha512-Cxwpt2SfTzTtXcfOlzGEee8O+c+MmUgGrNiBcXnuWxuFJHe6a5Hz7qwhwe5OgaSYI0IJvkLqWX1ASG+cJOkEiA=="],
 
     "pretty-format/react-is": ["react-is@18.3.1", "", {}, "sha512-/LLMVyas0ljjAtoYiPqYiL8VWXzUUdThrmU5+n20DZv+a+ClRoevUzw5JxU+Ieh5/c87ytoTBV9G1FiKfNJdmg=="],
@@ -1947,6 +1946,8 @@
 
     "wrap-ansi/strip-ansi": ["strip-ansi@6.0.1", "", { "dependencies": { "ansi-regex": "^5.0.1" } }, "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A=="],
 
+    "xml2js/xmlbuilder": ["xmlbuilder@11.0.1", "", {}, "sha512-fDlsI/kFEx7gLvbecc0/ohLG50fugQp8ryHzMTuW9vSa1GJ0XYWKnhsUx7oie3G98+r56aTQIUB4kht42R3JvA=="],
+
     "@expo/fingerprint/minimatch/brace-expansion": ["brace-expansion@5.0.6", "", { "dependencies": { "balanced-match": "^4.0.2" } }, "sha512-kLpxurY4Z4r9sgMsyG0Z9uzsBlgiU/EFKhj/h91/8yHu0edo7XuixOIH3VcJ8kkxs6/jPzoI6U9Vj3WqbMQ94g=="],
 
     "@expo/metro-config/hermes-parser/hermes-estree": ["hermes-estree@0.33.3", "", {}, "sha512-6kzYZHCk8Fy1Uc+t3HGYyJn3OL4aeqKLTyina4UFtWl8I0kSL7OmKThaiX+Uh2f8nGw3mo4Ifxg0M5Zk3/Oeqg=="],

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",

src/app/(tabs)/settings.tsx

@@ -6,6 +6,8 @@ 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';
 
@@ -19,6 +21,24 @@ export default function SettingsScreen() {
   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}>
@@ -40,7 +60,7 @@ export default function SettingsScreen() {
                   {selected && <ThemedText type="small" style={{ color: ACCENT }}>✓ selected</ThemedText>}
                 </View>
                 <ThemedText type="small" themeColor="textSecondary">
-                  {cap(m.tier)} · ~{m.approxMB} MB · {m.multilingual ? 'multilingual' : 'English-only'}
+                  {cap(m.tier)} · ~{Math.round(m.approxMB * sizeMult)} MB · {m.multilingual ? 'multilingual' : 'English-only'}
                 </ThemedText>
               </ThemedView>
             </Pressable>

src/lib/audio/decode.native.ts

@@ -24,7 +24,7 @@ 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';
 
@@ -133,7 +133,12 @@ export const decoder: AudioDecoder = {
     const arrayBuffer = new Uint8Array(bytes).buffer as ArrayBuffer;
     let audioBuffer;
     try {
-      audioBuffer = await decodeAudioData(arrayBuffer);
+      // 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(

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.native.ts

@@ -411,7 +411,14 @@ function getDb(): Promise<SQLite.SQLiteDatabase> {
       const db = await SQLite.openDatabaseAsync('wisp.db');
       await runMigrations(db);
       return db;
-    })();
+    })().catch((e) => {
+      // Don't cache a REJECTED promise — otherwise a transient open/migration
+      // failure (locked db, backgrounded mid-migration) would brick storage for
+      // the whole session, with every later call returning the same rejection.
+      // Clearing it lets the next getDb() retry from scratch.
+      dbPromise = undefined;
+      throw e;
+    });
   }
   return dbPromise;
 }
@@ -653,14 +660,17 @@ export const repo: StorageRepo = {
     const transcript = JSON.parse(row.json) as Transcript;
     const updatedAt = Date.now();
     const next: Transcript = { ...transcript, courseId, updatedAt };
-    await db.withTransactionAsync(async () => {
-      await db.runAsync(
+    // Exclusive so the two writes commit atomically and can't interleave with
+    // concurrent repo queries (withTransactionAsync is NOT isolated on a shared
+    // connection — see expo-sqlite docs).
+    await db.withExclusiveTransactionAsync(async (txn) => {
+      await txn.runAsync(
         `UPDATE transcripts SET courseId = ?, updatedAt = ?, json = ? WHERE id = ?`,
         [courseId, updatedAt, JSON.stringify(next), transcriptId],
       );
       // Keep the denormalized courseId on every vector row in sync so
       // course-scoped search keeps matching after a move.
-      await db.runAsync(`UPDATE segvecs SET courseId = ? WHERE transcriptId = ?`, [
+      await txn.runAsync(`UPDATE segvecs SET courseId = ? WHERE transcriptId = ?`, [
         courseId,
         transcriptId,
       ]);
@@ -763,22 +773,22 @@ export const repo: StorageRepo = {
     const db = await getDb();
     // First reassign this course's transcripts to Unsorted, THEN delete the
     // course row, so we never leave dangling courseId references.
-    await db.withTransactionAsync(async () => {
-      await db.runAsync(
+    await db.withExclusiveTransactionAsync(async (txn) => {
+      await txn.runAsync(
         `UPDATE transcripts SET courseId = NULL WHERE courseId = ?`,
         [id],
       );
       // Keep denormalized vector rows consistent: their transcripts are now
       // Unsorted, so their courseId must follow.
-      await db.runAsync(`UPDATE segvecs SET courseId = NULL WHERE courseId = ?`, [
+      await txn.runAsync(`UPDATE segvecs SET courseId = NULL WHERE courseId = ?`, [
         id,
       ]);
       // Flashcards in this course follow their transcripts to Unsorted.
-      await db.runAsync(
+      await txn.runAsync(
         `UPDATE flashcards SET courseId = NULL WHERE courseId = ?`,
         [id],
       );
-      await db.runAsync(`DELETE FROM courses WHERE id = ?`, [id]);
+      await txn.runAsync(`DELETE FROM courses WHERE id = ?`, [id]);
     });
   },
 
@@ -867,14 +877,18 @@ export const repo: StorageRepo = {
       [transcriptId],
     );
     const courseId = owner ? owner.courseId : null;
-    await db.withTransactionAsync(async () => {
+    // Exclusive: the DELETE-then-INSERT must be atomic + isolated, else a
+    // concurrent searchVectors during an index rebuild can read zero vectors,
+    // and a crash between the DELETE and the INSERTs would lose this
+    // transcript's embeddings entirely.
+    await db.withExclusiveTransactionAsync(async (txn) => {
       // Replace, not merge: clear existing rows so a re-embed with fewer
       // segments can't leave stale rows behind.
-      await db.runAsync(`DELETE FROM segvecs WHERE transcriptId = ?`, [
+      await txn.runAsync(`DELETE FROM segvecs WHERE transcriptId = ?`, [
         transcriptId,
       ]);
       for (const v of vectors) {
-        await db.runAsync(
+        await txn.runAsync(
           `INSERT INTO segvecs
              (transcriptId, segmentId, start, end, courseId, text, vector, model)
            VALUES (?, ?, ?, ?, ?, ?, ?, ?)`,
@@ -986,9 +1000,9 @@ export const repo: StorageRepo = {
         srs,
       };
     });
-    await db.withTransactionAsync(async () => {
+    await db.withExclusiveTransactionAsync(async (txn) => {
       for (const card of cards) {
-        await db.runAsync(
+        await txn.runAsync(
           `INSERT INTO flashcards
              (id, transcriptId, courseId, segmentId, start, front, back,
               createdAt, srs, dueAt)

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

src/lib/enrich/bib.ts

@@ -4,12 +4,28 @@ import type { RefMetadata } from './types';
 
 /** Serialize references to BibTeX. Handles missing fields and []. */
 export function toBibtex(refs: RefMetadata[]): string {
-  return refs.map((r, i) => bibEntry(r, i)).join('\n\n');
+  // 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, index: number): string {
+function bibEntry(r: RefMetadata, key: string): string {
   const type = r.authors && r.authors.length ? 'article' : 'misc';
-  const key = bibKey(r, index);
 
   const fields: string[] = [];
   if (r.title) fields.push(`  title = {${brace(r.title)}}`);

src/lib/enrich/dates.ts

@@ -159,6 +159,20 @@ function detectIso(text: string, anchorMs: number): RawMatch[] {
   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[] = [];
@@ -175,6 +189,9 @@ function detectMonthName(text: string, anchorMs: number): RawMatch[] {
     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);
   }
 
@@ -188,6 +205,9 @@ function detectMonthName(text: string, anchorMs: number): RawMatch[] {
     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);
   }
 

src/lib/learn/glossary.ts

@@ -103,9 +103,11 @@ export function glossary(
       const m = sentence.match(DEFINITION_RE);
       if (!m || !m[1]) continue;
       const term = m[1].trim();
-      // Skip junk subjects (pure stopword / too short).
+      // 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;
+      if (termWords.length === 0 || term.length < 3) continue;
       add({
         term,
         definition: sentence,

src/lib/learn/srs.test.ts

@@ -79,4 +79,17 @@ describe('review', () => {
     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

@@ -60,10 +60,14 @@ export function review(
     lapses += 1;
     intervalDays = AGAIN_INTERVAL_DAYS;
   } else if (grade === 1) {
-    // Hard: passed, but gently grow the interval and lower ease.
+    // 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 = nextInterval(srs, ease, HARD_MULTIPLIER);
+    intervalDays = hardInterval(srs);
   } else if (grade === 2) {
     // Good: standard SM-2 progression, ease unchanged.
     reps += 1;
@@ -104,6 +108,16 @@ function nextInterval(prev: SrsState, ease: number, multiplier: number): number
   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/transcription/engineImpl.web.ts

@@ -156,13 +156,26 @@ export const engine: TranscriptionEngine = {
 
     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 ?? []) {
-      const [start, end] = c.timestamp;
+    for (let i = 0; i < chunks.length; i++) {
+      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.ts

@@ -108,6 +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.

src/stores/transcriptsStore.ts

@@ -7,6 +7,11 @@ 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;
@@ -27,6 +32,7 @@ export const useTranscripts = create<TranscriptsState>((set, get) => ({
   courseFilter: 'all',
 
   refresh: async () => {
+    const my = ++refreshToken;
     set({ loading: true });
     try {
       const repo = getRepo();
@@ -40,9 +46,13 @@ export const useTranscripts = create<TranscriptsState>((set, get) => ({
       } 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 });
     }
   },