Build Wisp: on-device transcription studio (web + native, one codebase)

Private, offline speech-to-text that runs Whisper on the user's own device —
free, no account, no per-minute fees. Replaces Otter.ai / Rev.

- Pure, tested engine: chunking, overlap timestamp-stitching, exports
  (SRT/VTT/TXT/MD/JSON), WAV codec, resampler, job queue, model catalog (142 tests).
- Platform-abstracted TranscriptionEngine: transformers.js on web (loaded from
  CDN at runtime to dodge Metro's onnxruntime-web bundling limits), whisper.rn
  on native. Shared pipeline orchestrates decode -> chunk -> transcribe -> stitch.
- Cross-platform StorageRepo (Dexie web / expo-sqlite native), Zod-validated.
- UI: library + search, import, live-progress transcription, synced click-to-seek
  editor, multi-format export; model picker + privacy in settings.
- Web ships as a single-page PWA with COOP/COEP isolation for threaded WASM;
  Docker (nginx) image + Traefik compose for wisp.briggen.dev.

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

src/lib/transcription/engine.ts

@@ -0,0 +1,60 @@
+// The transcription engine interface: the single abstraction that hides the
+// platform-specific Whisper backend (transformers.js on web, whisper.rn on
+// native) behind one stable contract. The pure orchestration pipeline
+// (pipeline.ts) is written entirely against this interface so it can be unit
+// tested with a fake engine and never imports any heavy/native dependency.
+//
+// IMPORTANT contract detail: `transcribeChunk` returns CHUNK-LOCAL times
+// (0-based, relative to the start of the chunk it was handed). The pipeline is
+// responsible for offsetting those into absolute media time and stitching the
+// overlapping windows together (see pipeline.ts + stitch.ts).
+
+import type {
+  Backend,
+  ModelId,
+  PcmAudio,
+  Segment,
+  TranscribeOptions,
+} from '../types';
+
+/**
+ * What an engine discovered about the device it is running on. Used by the UI
+ * to gate model choices and to decide whether live-mic capture is available.
+ */
+export interface EngineCapabilities {
+  /** The compute backend the engine resolved to (e.g. 'webgpu', 'cpu'). */
+  backend: Backend;
+  /** Whether this engine can transcribe from a live microphone stream. */
+  supportsLiveMic: boolean;
+  /** Largest model we'd recommend running on this device/backend. */
+  maxRecommendedModel: ModelId;
+}
+
+/**
+ * A platform-agnostic Whisper engine. Both `engineImpl.web.ts` and
+ * `engineImpl.native.ts` export a singleton `engine` satisfying this.
+ */
+export interface TranscriptionEngine {
+  /** Which platform family this engine targets. */
+  readonly platform: 'web' | 'native';
+
+  /** Probe the device. May be async (e.g. WebGPU adapter request). */
+  capabilities(): Promise<EngineCapabilities>;
+
+  /**
+   * Ensure `modelId` is loaded and ready. `onProgress` (if given) receives a
+   * fraction in [0, 1] during download/initialization. Implementations should
+   * be idempotent: calling again for an already-loaded model is a no-op.
+   */
+  loadModel(modelId: ModelId, onProgress?: (p: number) => void): Promise<void>;
+
+  /** Synchronous check: is this model already loaded in memory? */
+  isModelLoaded(modelId: ModelId): boolean;
+
+  /**
+   * Transcribe a single already-decoded 16 kHz mono chunk.
+   * Returns segments with CHUNK-LOCAL times (seconds, 0-based). The caller
+   * offsets and stitches; engines must NOT add the chunk's media offset.
+   */
+  transcribeChunk(audio: PcmAudio, opts: TranscribeOptions): Promise<Segment[]>;
+}