Renumber spec 001 decision placeholders to D-040–D-051; record the Danish publication

Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-10-30 01:47:27 +01:00
parent 30969e86b7
commit 0632f7843b
71 changed files with 244 additions and 236 deletions
@@ -40,12 +40,12 @@ step with the commands above, in the same commit.
  `Bun.$`. Bun API docs: `node_modules/bun-types/docs/`.
 - **Two engines, two libraries (D-013).** Browser inference = Transformers.js, inside
  `src/worker/inference.worker.ts` only (transcript translation: `src/worker/translate.worker.ts`,
-  D-WS4b) — never on the main thread. Server inference = faster-whisper
+  D-045) — never on the main thread. Server inference = faster-whisper
  via `server-py/sidecar.py` — never Transformers.js on the server. The one exception is Parakeet
-  (`family: "parakeet"`, D-WS7a): parakeet.js in the same worker, on the **same** onnxruntime-web
+  (`family: "parakeet"`, D-051): parakeet.js in the same worker, on the **same** onnxruntime-web
  instance — never call its `fromUrls`/`fromHub`/`initOrt` (own ort copy, CDN wasm path).
 - **`@huggingface/transformers` is patched and a devDependency on purpose.** It is pinned by
-  `patches/@huggingface%[email protected]` (D-027/D-028, D-WS1a) and kept out of runtime deps so the
+  `patches/@huggingface%[email protected]` (D-027/D-028, D-040) and kept out of runtime deps so the
  Docker image stays small (D-031). Upgrading or moving it means redoing the patch: ask first.
 - **No cross-origin subresources.** Every response carries COOP/COEP `require-corp` (D-023,
  `src/index.ts`); a CDN script/style/font will be blocked. Self-host it (onnxruntime is served from `/ort/`).
@@ -13,7 +13,7 @@ On desktop browsers the on-device engine also offers **NVIDIA Parakeet TDT 0.6B
 same onnxruntime-web. Model: [nvidia/parakeet-tdt-0.6b-v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3),
 © NVIDIA, licensed [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/); ONNX
 conversion [ysdede/parakeet-tdt-0.6b-v3-onnx](https://huggingface.co/ysdede/parakeet-tdt-0.6b-v3-onnx),
-downloaded unmodified by the browser on first use (Decisions.md D-WS7a).
+downloaded unmodified by the browser on first use (Decisions.md D-051).

 See `docs/Architecture.md` for the design, `docs/Decisions.md` for the decision
 log, and `PRODUCT.md` for the product spec.
@@ -27,7 +27,7 @@ Logged-in users may optionally route inference to the server.
 │   ├─ Device selection: WebNN → WebGPU → WASM fallback            │
 │   └─ Model weights cached via browser Cache API                  │
 │                                                                  │
-│  Translate Web Worker (D-WS4b)                                   │
+│  Translate Web Worker (D-045)                                   │
 │   └─ Transformers.js Opus-MT (MarianMT), WASM, on user request   │
 │                                                                  │
 │  Service Worker (PWA)                                            │
@@ -76,19 +76,19 @@ src/
    parakeet.ts            # pure Parakeet logic: ≤30 s window plan, words→chunks, vocab tokenizer, pinned URLs
    modelCache.ts          # "downloaded?" per model via Cache Storage (D-005; reused by M3 offline indicator)
    engine.ts              # engine abstraction: BrowserEngine | ServerEngine
-    translation.ts         # transcript translation planning: Opus-MT catalog, units, sentence pieces (pure, D-WS4b)
-    hint.ts                # "Names and terms" hint: normalize, localStorage, prompt truncation rule (D-WS2a)
+    translation.ts         # transcript translation planning: Opus-MT catalog, units, sentence pieces (pure, D-045)
+    hint.ts                # "Names and terms" hint: normalize, localStorage, prompt truncation rule (D-048)
  worker/
    inference.worker.ts    # Transformers.js pipeline, WebNN/WebGPU/WASM selection; dispatches on catalog `family`
-    parakeetEngine.ts      # Parakeet TDT path: parakeet.js ParakeetModel on the shared onnxruntime-web (D-WS7a)
+    parakeetEngine.ts      # Parakeet TDT path: parakeet.js ParakeetModel on the shared onnxruntime-web (D-051)
    modelFiles.ts          # Cache-API weight download for non-Transformers.js paths (same cache, same progress msgs)
    ortShim.ts             # re-exports onnxruntime-web/webgpu; parakeet.js's own ort import is aliased here
    protocol.ts            # typed postMessage messages (shared with UI)
-    prompt.ts              # decoder prompt (<|startofprev|> hint) via a generate() wrapper (D-WS2a)
-    translate.worker.ts    # transcript translation (Opus-MT), its own worker + ort instance (D-WS4b)
+    prompt.ts              # decoder prompt (<|startofprev|> hint) via a generate() wrapper (D-048)
+    translate.worker.ts    # transcript translation (Opus-MT), its own worker + ort instance (D-045)
    translateProtocol.ts   # typed messages for the translate worker
    opusMt.ts              # Opus-MT loader shared by the worker, bench script and real-model test
-    longform/              # long-form file pipeline (D-WS3a): pure vad.ts / quality.ts / pipeline.ts
+    longform/              # long-form file pipeline (D-049): pure vad.ts / quality.ts / pipeline.ts
                           # behind transcribeLongForm(audio, deps, options); transformers.ts implements
                           # the deps (Silero VAD + Whisper decodes) on the same Transformers.js runtime
  pwa/
@@ -158,17 +158,17 @@ environment). See `WorkBreakdown.md` for the itemized checklist.)
   `AudioContext.decodeAudioData`, downmixes to mono, resamples to 16 kHz,
   transfers the `Float32Array` to the worker. For multilingual models the
   `transcribe` message also carries `language` (or auto-detect) and `task`
-   (transcribe | translate — translate only for models with `canTranslate`, D-WS4a).
+   (transcribe | translate — translate only for models with `canTranslate`, D-044).
   For every model it may carry `prompt`: the optional "Names and terms" hint
-   (`VocabularyHint`, stored in localStorage by `lib/hint.ts`), raw as typed (D-WS2a).
+   (`VocabularyHint`, stored in localStorage by `lib/hint.ts`), raw as typed (D-048).
 5. Worker runs the pipeline (chunked, with timestamps), streaming partial results
   back via `{type:"partial"}` messages; final result as `{type:"result"}`.
   A hint becomes `<|startofprev|> hint.` once per pass (`worker/prompt.ts`,
   cut to a 120-token prompt budget at a whole word) and is prepended
   to the decoder input of every 30 s window by a wrapper around the model's
-   `generate()`, which also strips it from the output again (D-WS2a).
+   `generate()`, which also strips it from the output again (D-048).
   Live passes (4.1b) carry the same `prompt`.
-6. **Long-form branch (D-WS3a/D-WS3b, "Advanced → Skip silence", off by default).**
+6. **Long-form branch (D-049/D-050, "Advanced → Skip silence", off by default).**
   With the option on, every load posted for a file run or the Download button carries
   `vad: true`; once the Whisper pipeline is ready the worker loads Silero VAD
   (`onnx-community/silero-vad`, 2.2 MB, Cache API like the weights; offline it is
@@ -183,7 +183,7 @@ environment). See `WorkBreakdown.md` for the itemized checklist.)
   chunked path whenever the VAD model isn't loaded, and live passes never take
   this branch (live keeps its RMS segmenter, D-016/D-024). A finished transcript
   with no words shows "No speech was found in this recording." The names-and-terms
-   hint (`prompt`) is applied to every chunk's decode the same way (D-WS2a).
+   hint (`prompt`) is applied to every chunk's decode the same way (D-048).

 ### 4.1b Live microphone transcription (anonymous, browser engine — D-016)
 1. `LiveControls` starts a session: `getUserMedia` → `AudioContext` at the
@@ -216,12 +216,12 @@ environment). See `WorkBreakdown.md` for the itemized checklist.)
   only when signed in with the server engine available (`EngineToggle`);
   anonymous/offline users always get the browser-engine path above.

-### 4.1c Transcript translation (any engine — D-WS4b)
+### 4.1c Transcript translation (any engine — D-045)

 1. Any finished transcript (browser file, browser live, server engine) shows a
   "Translate transcript…" disclosure (`TranslatePanel`). Whisper's own
   "Translate to English" task stays in `TranscriptionOptions`, only for
-   models with `canTranslate` (D-WS4a).
+   models with `canTranslate` (D-044).
 2. On **Translate**, the UI groups the transcript's chunks into sentence-ending
   units (`buildTranslationUnits`, merged time spans) and posts their texts to
   `translate.worker.js` — a separate worker with its own onnxruntime instance.
@@ -248,7 +248,7 @@ environment). See `WorkBreakdown.md` for the itemized checklist.)
 - Model weights live in the Cache API (`transformers-cache`) via Transformers.js —
  independent of the SW, which never touches that cache (D-005). `/api/*` is
  network-only. The Silero VAD model of the long-form branch lives there too
-  (D-WS3b); offline, the worker checks that cache before loading it and never
+  (D-050); offline, the worker checks that cache before loading it and never
  fetches, falling back to the chunked path if it is missing.
 - Updates: `pwa/register.ts` surfaces a waiting SW as a toast; accepting posts
  `SKIP_WAITING`, activation drops old `whisper-*` caches, and the page reloads.
@@ -266,11 +266,11 @@ environment). See `WorkBreakdown.md` for the itemized checklist.)
   `src/server/transcribe.ts` bridges the sidecar's segment stream to the streamed
   JSON-lines HTTP response (D-011). `language`, `task` and the vocabulary hint
   (`prompt`) travel as query params (also on the live WebSocket) and are
-   validated by `parseLanguageAndTask` (D-032, D-WS2a); the sidecar cuts the
+   validated by `parseLanguageAndTask` (D-032, D-048); the sidecar cuts the
   hint with the same rule as the browser and passes it to faster-whisper as
-   `initial_prompt` (D-WS2a). File requests (`mode: "file"`) decode with
+   `initial_prompt` (D-048). File requests (`mode: "file"`) decode with
   faster-whisper's batched pipeline over Silero-VAD speech chunks; server-live
-   passes send `mode: "live"` and stay sequential (D-WS5a). The shared catalog maps model ids to
+   passes send `mode: "live"` and stay sequential (D-047). The shared catalog maps model ids to
   `Systran/faster-whisper-*` CTranslate2 conversions. If the sidecar is absent,
   the server engine reports unavailable; the browser/anonymous flow is unaffected.
 4. Both engines implement the same `Engine` interface in `src/lib/engine.ts`, so the
@@ -283,7 +283,7 @@ environment). See `WorkBreakdown.md` for the itemized checklist.)
 - Anonymous users are never blocked from browser inference — auth only unlocks the
  server engine.

-### 4.1c Parakeet (English-only, desktop-only browser model — D-WS7a)
+### 4.1c Parakeet (English-only, desktop-only browser model — D-051)
 Catalog entries carry `family?: "whisper" | "parakeet"` (absent ⇒ whisper). For
 `family: "parakeet"` the worker skips the Transformers.js pipeline:
 1. **Load:** `modelFiles.ts` fetches `vocab.txt`, the decoder_joint and the
@@ -316,7 +316,7 @@ Catalog entries carry `family?: "whisper" | "parakeet"` (absent ⇒ whisper). Fo
 - **Anonymous-first:** every feature except the server engine works logged out.
 - **Offline-first:** app shell + previously-downloaded models must work offline.
 - **One browser inference runtime:** onnxruntime-web, driven by Transformers.js for
-  Whisper and by parakeet.js for Parakeet (D-WS7a) — the same ort instance, never a
+  Whisper and by parakeet.js for Parakeet (D-051) — the same ort instance, never a
  second runtime; the server engine runs faster-whisper (D-013). The model catalog
  stays shared across both engines (browser-only entries have no faster-whisper name).
 - **Worker isolation:** inference never runs on the main thread.
@@ -125,7 +125,7 @@ than ONNX Runtime under Bun; the server engine becomes genuinely useful rather t
 slow convenience.

 ## D-014 — WASM sessions run at graphOptimizationLevel "basic"
-**Status:** **Superseded by D-WS1c** (bug workaround, 2026-07-10; the ort-web
+**Status:** **Superseded by D-043** (bug workaround, 2026-07-10; the ort-web
 build in Transformers.js 4.3.2 no longer has the bug, so the cap is gone)
 On the WASM backend, the inference worker passes
 `session_options: { graphOptimizationLevel: "basic" }` to the Transformers.js
@@ -1226,7 +1226,7 @@ exists to catch.
 **Consequence:** full `bun test` ≈ 19 min on CPU (en_alien ≈ 20 min of it
 on a cold run); the fast suite excludes this file.

-## D-WS1a — Transformers.js 4.2.0 → 4.3.2, sequential-session patch carried over
+## D-040 — Transformers.js 4.2.0 → 4.3.2, sequential-session patch carried over
 **Status:** Accepted (2026-10-29, spec 001 Goal 1)
 `@huggingface/transformers` is now pinned to exactly `4.3.2` (latest; 4.3.0
 added Safari 26+ WebGPU). It brings onnxruntime-web
@@ -1254,7 +1254,7 @@ TDT/RNN-T (transducer) decoder or joint-network loop anywhere in the
 package, so **Parakeet TDT 0.6B v3 is not supported natively**; per the
 spec, that path uses `parakeet.js`.

-## D-WS1d — WebGPU-side effects of the 4.3.2 upgrade: inspected, not measured (needs the owner's browser check)
+## D-041 — WebGPU-side effects of the 4.3.2 upgrade: inspected, not measured (needs the owner's browser check)
 **Status:** Open — inspection only (2026-10-29)
 This machine has no GPU and no browser WebGPU, so D-022, D-025 and D-026
 cannot be re-measured here. What the upgrade changes on that side, from the
@@ -1294,7 +1294,7 @@ cannot be re-measured here. What the upgrade changes on that side, from the
 5. **Safari 26 (macOS/iOS):** what the footer reports for "auto"; if WebGPU,
   check speed and output.

-## D-WS1b — D-021's `max_new_tokens: 440` is still needed on Transformers.js 4.3.2
+## D-042 — D-021's `max_new_tokens: 440` is still needed on Transformers.js 4.3.2
 **Status:** Accepted (2026-10-29) — re-test of D-021 part 3, which stands unchanged
 **Source:** 4.3.2's `WhisperForConditionalGeneration.generate()` still takes
 the seek loop on exactly D-021's condition,
@@ -1321,10 +1321,10 @@ version-to-version speed comparison.
 **Decision:** keep `max_new_tokens: 440` in `src/worker/inference.worker.ts`
 and in `tests/live-real-model.test.ts`.

-## D-WS1c — WASM sessions back at the default graph optimization level (D-014's "basic" cap dropped)
+## D-043 — WASM sessions back at the default graph optimization level (D-014's "basic" cap dropped)
 **Status:** Accepted (2026-10-30) — supersedes D-014
 **Re-test** of D-014's revisit condition after the Transformers.js 4.3.2 bump
-(D-WS1a), which brings onnxruntime-web `1.31.0-dev.20260914-8d85527a0`.
+(D-040), which brings onnxruntime-web `1.31.0-dev.20260914-8d85527a0`.
 Measured under Bun with onnxruntime-web's own wasm EP (not onnxruntime-node),
 1 thread, AMD Ryzen 9 9950X shared with other workstreams.
 1. **Session creation**, `onnx-community/whisper-tiny`
@@ -1361,7 +1361,7 @@ check is re-run.
 `ort-wasm-simd-threaded.*` files, but multithreaded under COOP/COEP) —
 `tests/offline.md` §2's WASM step is the owner's browser check.

-## D-WS4a — Whisper's "Translate to English" only on models trained for it (`canTranslate`)
+## D-044 — Whisper's "Translate to English" only on models trained for it (`canTranslate`)
 **Status:** Accepted (2026-10-29) · spec 001 Goal 4a
 `WhisperModel` gains `canTranslate: boolean` (`src/lib/models.ts`): **true** for multilingual
 tiny / base / small / medium, **false** for every `.en` model (no task token) and for
@@ -1376,13 +1376,13 @@ server's `parseLanguageAndTask` (D-032) still accepts "translate" for any model
 sends it for turbo now, and a direct API caller asking faster-whisper to translate with turbo gets
 what faster-whisper does; rejecting it server-side would be a behaviour change for API users with no
 UI benefit. Speech in another language than English can still end up in English (or Danish) through
-text translation of the finished transcript (D-WS4b), for every model.
+text translation of the finished transcript (D-045), for every model.

-## D-WS4b — Text translation of finished transcripts with Opus-MT, in its own worker
+## D-045 — Text translation of finished transcripts with Opus-MT, in its own worker
 **Status:** Accepted (2026-10-29) · spec 001 Goal 4b
 **What:** any finished transcript — browser file, browser live, server engine — can be translated
 da→en or en→da from a secondary "Translate transcript…" disclosure under the transcript
-(`TranslatePanel`). Whisper's own task keeps its label "Translate to English" (D-WS4a). The
+(`TranslatePanel`). Whisper's own task keeps its label "Translate to English" (D-044). The
 translation renders as a second transcript card ("Translation (English)") with its own
 `.txt` / `.srt` / `.txt+ts` exports named `<file>.<en|da>.<ext>`.
 **Models:** `Xenova/opus-mt-da-en` and `Xenova/opus-mt-en-da` — Transformers.js ONNX exports of
@@ -1432,7 +1432,7 @@ terminates the translate worker only.
 **Not done / follow-ups:** WebGPU for Opus-MT (unmeasurable here); more language pairs (the catalog
 is two entries, `TRANSLATION_MODELS`); a translation of live text *while* recording.

-## D-WS6a — Danish fine-tuned Whisper models: our own ONNX + CTranslate2 conversions, unpublished until the owner says go
+## D-046 — Danish fine-tuned Whisper models: our own ONNX + CTranslate2 conversions; only whisper-small-danish published
 **Status:** Proposed (2026-10-29, spec 001 Goal 6)
 **Candidates checked (licence + architecture, from each repo's model card, config and notices):**
 - `thorhojhus/whisper-small-danish` — MIT; plain `WhisperForConditionalGeneration`, whisper-small
@@ -1484,7 +1484,15 @@ use beam 5.
 cards credit the original authors and carry MIT, or Apache-2.0 + NOTICE), then set
 `DANISH_MODELS_HF_OWNER`. The integrator sets `canTranslate: false` on these entries at merge.

-## D-WS5a — Server engine: batched faster-whisper + tuned Silero VAD for files; live passes stay sequential
+**Publication (2026-10-30, owner's decision):** only the small fine-tune is
+published — `lucasw89/whisper-small-danish-ONNX` (browser) and
+`lucasw89/whisper-small-danish-ct2-int8` (server), public, model cards
+crediting the `thorhojhus` account and carrying MIT. The turbo-danish and
+edda-v0.2 conversions showed no clear gain over stock turbo on `da_festuge`, so
+they stay unpublished (folders kept locally in `/home/user/hf-publish/`) and
+their catalog entries were removed. `DANISH_MODELS_HF_OWNER = "lucasw89"`.
+
+## D-047 — Server engine: batched faster-whisper + tuned Silero VAD for files; live passes stay sequential
 **Status:** Accepted (2026-10-30, spec 001 Goal 5)
 `server-py/sidecar.py` decoded every request with the sequential
 `WhisperModel.transcribe`, no VAD.
@@ -1529,7 +1537,7 @@ long. Segments arrive in bursts of 8 chunks (≈ 4 min of speech) and the first
 burst can take ~15–25 s on a long file; the D-021 heartbeat covers it, and
 cancel lands at the next burst boundary. No new env vars.

-## D-WS2a — "Names and terms" vocabulary hint as a Whisper decoder prompt, cut to 120 tokens
+## D-048 — "Names and terms" vocabulary hint as a Whisper decoder prompt, cut to 120 tokens
 **Status:** Accepted (2026-10-29) — spec 001, Goal 2. Placeholder id; the
 integrator renumbers it.
 **Context:** users want names and jargon spelled their way. Whisper takes a
@@ -1613,14 +1621,14 @@ test; revisit if a larger list is wanted on turbo).
 long hint. WebGPU behaviour of the prompt path is unmeasured here (CPU
 only) — needs the owner's browser check.

-**Integration note (2026-10-30, merge with D-WS5a):** faster-whisper's
+**Integration note (2026-10-30, merge with D-047):** faster-whisper's
 `BatchedInferencePipeline` (server file mode) re-encodes `initial_prompt`
 itself and accepts only text, so the sidecar passes it the budget-cut ids
 decoded back to text (re-encoding gives the same ids, checked with the real
 tokenizer incl. "Århus, Ærø"); server-live passes keep the ids. Pinned by the
 stubbed-sidecar test "the vocabulary hint reaches the batched pipeline as text".

-## D-WS3a — Long-form file pipeline (Silero VAD chunks, language pinned once) ships as an Advanced option, off by default
+## D-049 — Long-form file pipeline (Silero VAD chunks, language pinned once) ships as an Advanced option, off by default
 **Status:** Accepted (2026-10-30) — spec 001 Goal 3
 Browser file transcription cuts audio into fixed 30 s windows with 5 s
 overlap (Transformers.js's chunked pipeline), decoding silence and
@@ -1678,14 +1686,14 @@ bars. Whisper's `generate()` drops kwargs not in `forward_params`, so
 instance; a Transformers.js upgrade must keep the "one encoder pass per
 chunk" assertion green. WebGPU timings are unmeasured.

-**Integration note (2026-10-30, merge with D-WS2a):** the long-form adapter
+**Integration note (2026-10-30, merge with D-048):** the long-form adapter
 (`worker/longform/transformers.ts`) applies the names-and-terms hint to every
 chunk's decode with the same decoder-prompt prefix as the chunked path
 (tokenized once per pass); before the merge the `prompt` option reached the
 adapter but was ignored. Pinned by a unit test in
 `tests/longform-transformers.test.ts`.

-## D-WS3b — Silero VAD model: onnx-community/silero-vad from the Hub, cached like Whisper weights, never fetched offline
+## D-050 — Silero VAD model: onnx-community/silero-vad from the Hub, cached like Whisper weights, never fetched offline
 **Status:** Accepted (2026-10-30) — spec 001 Goal 3
 **Decision:** the long-form pipeline's VAD is Silero VAD v5 as exported by
 `onnx-community/silero-vad` (`onnx/model.onnx`, fp32, 2 243 022 bytes, MIT;
@@ -1704,7 +1712,7 @@ path (`choosePipeline`). Live mode keeps its energy gate (D-016).
 off-by-default option); running it on the main thread (D-013).
 **Consequence:** the first "Skip silence" run online downloads 2.2 MB.

-## D-WS7a — Parakeet TDT 0.6B v3 as an English-only, desktop-only browser model via parakeet.js on the shared onnxruntime-web
+## D-051 — Parakeet TDT 0.6B v3 as an English-only, desktop-only browser model via parakeet.js on the shared onnxruntime-web
 **Status:** Accepted (2026-10-29, spec 001 Goal 7)
 **Runtime decision (evidence).** `@huggingface/transformers` 4.2.0 (ours) and
 4.3.2 (npm `latest`) were inspected from their published tarballs: both ship
@@ -66,10 +66,10 @@ Notes (M1):
 - [x] Handle model switching (dispose old pipeline, load new)
 - [x] Language / task options for multilingual models (transcribe vs translate)
 - [x] Whisper translate only on models trained for it (`canTranslate`; hidden for
-      large-v3-turbo and `.en`, stale "translate" resets to transcribe — D-WS4a)
+      large-v3-turbo and `.en`, stale "translate" resets to transcribe — D-044)
 - [~] Verify quantized variants (q4/q8/fp16) on both backends; record working
      combinations in `models.ts`
- [~] Danish fine-tunes (spec 001 Goal 6, D-WS6a): `whisper-small-danish`,
+- [~] Danish fine-tunes (spec 001 Goal 6, D-046): `whisper-small-danish`,
      `whisper-large-v3-turbo-danish`, `edda-v0.2`. We converted them to ONNX
      (fp32/fp16/q8/q4) and CT2 int8 and verified them on CPU. The catalog
      entries hang off `DANISH_MODELS_HF_OWNER`. Still pending: the owner's go
@@ -154,7 +154,7 @@ Notes (M4):
      Bun (D-013 supersedes the Transformers.js-under-Bun plan); map the shared
      model catalog to faster-whisper model ids in `src/lib/models.ts`
 - [x] Batched inference + tuned Silero VAD for server-engine file
-      transcriptions, live passes sequential (spec 001 Goal 5, D-WS5a):
+      transcriptions, live passes sequential (spec 001 Goal 5, D-047):
      ≈ 2× faster on CPU int8 at equal-or-better WER; segments ≤ 30 s chunks
 - [x] Degrade gracefully when Python/the sidecar is absent (server engine
      unavailable; browser engine untouched — anonymous path deploys with Bun alone)
@@ -352,10 +352,10 @@ cached.

 ## M8 — Transcription quality and speed (spec 001)
 - [x] Transformers.js 4.2.0 → 4.3.2, sequential-session patch regenerated
-      for 4.3.2 and guarded by `tests/transformers-runtime.test.ts` (D-WS1a).
+      for 4.3.2 and guarded by `tests/transformers-runtime.test.ts` (D-040).
 - [x] WASM sessions back at the default graph optimization level: ort-web
-      1.31.0-dev loads the q8 decoders at every level (D-WS1c supersedes D-014).
- [x] Vocabulary hint "Names and terms" (D-WS2a) — optional field for every
+      1.31.0-dev loads the q8 decoders at every level (D-043 supersedes D-014).
+- [x] Vocabulary hint "Names and terms" (D-048) — optional field for every
      model, remembered in localStorage, sent as `prompt` on file and live
      passes (browser and server engines); cut to 120 tokens at a whole word.
      faster-whisper gets `initial_prompt` only (hotwords measured, no gain).
@@ -364,15 +364,15 @@ cached.

 ## Spec 001 — Translate (workstream 4)
 - [x] `canTranslate` per model; Task control hidden where false; stale
-      "translate" resets to transcribe (D-WS4a)
+      "translate" resets to transcribe (D-044)
 - [x] Text translation of any finished transcript, da→en / en→da, Opus-MT
-      in its own worker, sentence-chunked, timestamps kept for exports (D-WS4b)
+      in its own worker, sentence-chunked, timestamps kept for exports (D-045)
 - [~] Real-browser checks: offline translation (`tests/offline.md` §5),
      WebGPU-machine speed (owner)

 ---

-## Spec 001 Goal 7 — Parakeet TDT 0.6B v3 (English, desktop, D-WS7a)
+## Spec 001 Goal 7 — Parakeet TDT 0.6B v3 (English, desktop, D-051)
 - [x] Runtime decision: Transformers.js 4.2.0/4.3.2 have CTC-only Parakeet → parakeet.js on the
      shared onnxruntime-web (one runtime in the bundle, `tests/workerBundle.test.ts`).
 - [x] Catalog `family` + exact sizes; worker engine path (≤ 30 s windows, word → chunk
@@ -419,27 +419,27 @@ Found while bootstrapping agent docs (D-036); all fixed on 2026-10-29.
      — Light theme via tokens + self-hosted Inter (D-037). Also fixed: the
      Backend select overflowed its panel on 390 px-wide phones.
 - [ ] An English "Names and terms" hint can pull a short Danish live window
-      toward English words (one 2 s da_festuge snippet on turbo, D-WS2a).
+      toward English words (one 2 s da_festuge snippet on turbo, D-048).
      Not addressed: suggest same-language hints if it shows up in use.

-Found while building transcript translation (D-WS4b), 2026-10-29:
+Found while building transcript translation (D-045), 2026-10-29:
 - [ ] Transformers.js 4.2 batched Marian generation returns row 0's output for
      every row of a padded batch, so translation runs one sentence per call
      (~4× slower than batch 8 would be). Re-check after the 4.3 upgrade with
      `bun scripts/bench-translate.ts` ("output identical" across batch sizes).
 - [ ] The server API still accepts `task: "translate"` with large-v3-turbo
-      (only the UI hides it, D-WS4a).
+      (only the UI hides it, D-044).

-Found in spec 001 workstream 6 (D-WS6a), open:
+Found in spec 001 workstream 6 (D-046), open:
 - [ ] Transformers.js 4.2 has no beam search: `num_beams > 1` in a model's
      `generation_config.json` (edda-v0.2 ships 5) silently decodes greedily
-      and does a full-vocabulary top-k per step (D-WS6a). Not fixed: it would
+      and does a full-vocabulary top-k per step (D-046). Not fixed: it would
      mean overriding `num_beams: 1` in the worker, or a patch.
 - [ ] On `da_festuge`, neither Danish turbo fine-tune beats the stock turbo
-      (D-WS6a). Judging them needs a longer, harder Danish reference clip than
+      (D-046). Judging them needs a longer, harder Danish reference clip than
      this repo has.

-Found while building the long-form pipeline (spec 001 Goal 3, D-WS3a), 2026-10-29:
+Found while building the long-form pipeline (spec 001 Goal 3, D-049), 2026-10-29:
 - [ ] "Auto-detect" on the default (chunked) file path does not detect: with no
      `language`, Transformers.js 4.2 logs "No language specified - defaulting to
      English" and decodes Danish audio as an English translation (da_festuge WER
@@ -53,7 +53,7 @@ The integrator merges in this order, rebasing/fixing up each onto the result of
 - `AGENTS.md` rules apply in full (Bun only, COEP/no CDN, colour tokens, DI on server modules,
  `classifyError`, runtime-env config, TDD, docs duties).
 - **Decision ids:** don't take real `D-###` numbers. Write new entries as `## D-WS<n><letter> — …`
-  (e.g. `D-WS3a`) and cite the same placeholder in code comments; the integrator renumbers at merge.
+  (e.g. `D-049`) and cite the same placeholder in code comments; the integrator renumbers at merge.
 - **Heavy jobs take turns.** This machine has 16 cores / 18 GB RAM shared by all workstreams. Wrap
  every heavy command — real-model tests, the full `bun test`, benchmarks, model downloads,
  conversions, `pip install` — in `flock /tmp/whisper-heavy.lock <command>`. Light commands
@@ -1,5 +1,5 @@
 /**
- * Parakeet vs Whisper Large v3 Turbo benchmark (D-WS7a) — CPU only.
+ * Parakeet vs Whisper Large v3 Turbo benchmark (D-051) — CPU only.
 *
 *   flock /tmp/whisper-heavy.lock bun scripts/bench-parakeet.ts [parakeet-node|parakeet-web|turbo ...]
 *
@@ -1,5 +1,5 @@
 /**
- * Transcript-translation speed on onnxruntime-node (CPU), D-WS4b.
+ * Transcript-translation speed on onnxruntime-node (CPU), D-045.
 *
 *   flock /tmp/whisper-heavy.lock bun scripts/bench-translate.ts [batchSizes…]
 *
@@ -52,12 +52,12 @@ const workerBuild = await Bun.build({
  target: "browser",
  format: "esm",
  minify: true,
-  // One onnxruntime in the bundle (D-WS7a).
+  // One onnxruntime in the bundle (D-051).
  plugins: [ortAliasPlugin],
 });
 const workerScript = await workerBuild.outputs[0]!.text();
 await Bun.write(join(DIST, "inference.worker.js"), workerScript);
-// Transcript-translation worker (D-WS4b), same mechanism.
+// Transcript-translation worker (D-045), same mechanism.
 const translateWorkerBuild = await Bun.build({
  entrypoints: [join(ROOT, "src/worker/translate.worker.ts")],
  target: "browser",
@@ -1,5 +1,5 @@
 /**
- * Before/after measurement for the long-form pipeline (D-WS3a): wall-clock,
+ * Before/after measurement for the long-form pipeline (D-049): wall-clock,
 * WER against tests/audio_files/*.txt, generated tokens and decode calls, for
 * the worker's current chunked path ("old") and long-form configurations.
 *
@@ -1,7 +1,7 @@
 /**
 * Bun build plugin: parakeet.js's bare `onnxruntime-web` import resolves to
 * `src/worker/ortShim.ts` (= Transformers.js's `onnxruntime-web/webgpu`), so
- * the inference worker bundles exactly one onnxruntime (D-WS7a). Used by the
+ * the inference worker bundles exactly one onnxruntime (D-051). Used by the
 * production build (scripts/build.ts) and the dev server's worker bundle
 * (src/index.ts). Without it the bundle carries parakeet.js's own 1.24.1
 * copy as well — never executed, but ~0.5 MB of dead weight and a CDN
@@ -11,7 +11,7 @@ stdio — one JSON object per line, no framing beyond that:
             "audio_path": "/tmp/...", "language": "en" | null,
             "task": "transcribe" | "translate",
             "mode": "file" | "live",                 # optional, default "file"
-             "prompt": "Maxwell, Voices of the Void" | null}  # optional (D-WS2a)
+             "prompt": "Maxwell, Voices of the Void" | null}  # optional (D-048)
  stdout <- {"type": "segment", "id": "1", "start": 0.0, "end": 4.2,
             "text": " Hello world."}                       # 0..n of these
  stdout <- {"type": "done", "id": "1", "cancelled": false}
@@ -26,7 +26,7 @@ container decoding happens here. Requests are processed strictly one at a
 time; a reader thread keeps consuming stdin meanwhile so `cancel` can
 interrupt the segment loop of the request in flight.

-Decoding strategy by mode (Decisions.md D-WS5a):
+Decoding strategy by mode (Decisions.md D-047):
  "file"  faster-whisper's BatchedInferencePipeline with its Silero VAD
          (vad_filter=True, VAD_PARAMETERS): silence is dropped, speech is cut
          into <= 30 s chunks and BATCH_SIZE chunks decode per forward pass.
@@ -62,18 +62,18 @@ import sys
 import threading

 SAMPLE_RATE = 16000
-BATCH_SIZE = 8  # D-WS5a: VAD chunks decoded per batched forward pass
-# D-WS5a: Silero VAD tuned for file mode. faster-whisper's batched default
+BATCH_SIZE = 8  # D-047: VAD chunks decoded per batched forward pass
+# D-047: Silero VAD tuned for file mode. faster-whisper's batched default
 # (threshold 0.5, 400 ms padding) dropped quiet and overlapped speech —
 # deletions nearly doubled on en_alien; 0.2 / 800 ms brought WER back to the
 # sequential decoder's level. min_silence_duration_ms stays at VadOptions'
 # 2000 ms default (a dict replaces the pipeline's 160 ms preset).
 VAD_PARAMETERS = {"threshold": 0.2, "speech_pad_ms": 800}

-# "Names and terms" vocabulary hint (D-WS2a). Twins of normalizeHint /
+# "Names and terms" vocabulary hint (D-048). Twins of normalizeHint /
 # fitHintToBudget in src/lib/hint.ts — change both together; parity is tested
 # in tests/hint-parity.test.ts. Whisper would allow 223 tokens; 120 is the
-# measured cap at which small models are not garbled (D-WS2a).
+# measured cap at which small models are not garbled (D-048).
 PROMPT_TOKEN_BUDGET = 120
 _CONTROL_CHARS = re.compile("[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]")

@@ -215,7 +215,7 @@ def main():
            )
            if mode == "live":
                if prompt_ids:
-                    # D-WS2a: token ids, already cut to the budget by the
+                    # D-048: token ids, already cut to the budget by the
                    # same rule as the browser engine.
                    options["initial_prompt"] = prompt_ids
                segments, _info = cached_model.transcribe(
@@ -241,7 +241,7 @@ def main():
            # `segments` is a generator: inference happens during iteration,
            # so segments stream out (and cancel can cut in) while later ones
            # are still being decoded — per segment in live mode, per batch of
-            # BATCH_SIZE speech chunks in file mode (D-WS5a).
+            # BATCH_SIZE speech chunks in file mode (D-047).
            for segment in segments:
                if request_id in cancelled:
                    was_cancelled = True
@@ -80,10 +80,10 @@ interface AppErrorState {
 }

 const MODEL_IDS = MODELS.map((model) => model.id);
-// Downloaded badges cover the Opus-MT models too (transcript translation, D-WS4b).
+// Downloaded badges cover the Opus-MT models too (transcript translation, D-045).
 const CACHED_MODEL_IDS = [...MODEL_IDS, ...TRANSLATION_MODEL_IDS];

-/** Device signals for the mobile memory guard (D-027) and desktop-only models (D-WS7a). */
+/** Device signals for the mobile memory guard (D-027) and desktop-only models (D-051). */
 function memoryHint(): MobileMemoryHint {
  const nav = typeof navigator !== "undefined" ? navigator : undefined;
  return {
@@ -189,7 +189,7 @@ export function App() {
  const [activeThreads, setActiveThreads] = useState<number | null>(null);
  const [language, setLanguage] = useState(AUTO_LANGUAGE);
  const [task, setTask] = useState<WhisperTask>(DEFAULT_TASK);
-  // "Names and terms" vocabulary hint (D-WS2a): raw text, remembered in this
+  // "Names and terms" vocabulary hint (D-048): raw text, remembered in this
  // browser, sent with every pass for every model (`.en` included).
  const [hint, setHint] = useState(() => loadHint());
  const [downloaded, setDownloaded] = useState<ReadonlySet<string>>(new Set());
@@ -207,7 +207,7 @@ export function App() {
  const auth = useAuth();
  const [engineChoice, setEngineChoice] = useState<EngineKind>(() => loadEngineChoice());
  const [serverStatus, setServerStatus] = useState<ServerEngineStatus | null>(null);
-  // "Skip silence" (long-form pipeline, D-WS3a): remembered per browser. A
+  // "Skip silence" (long-form pipeline, D-049): remembered per browser. A
  // ref mirrors it for the worker-message handler set up once per worker.
  const [longForm, setLongForm] = useState(() => loadLongFormChoice());
  const longFormRef = useRef(longForm);
@@ -306,7 +306,7 @@ export function App() {
  function confirmRiskyMobileLoad(forModelId: string, preference: DevicePreference): boolean {
    const targetModel = getModel(forModelId);
    if (!targetModel) return true;
-    // Desktop-only models are never offered on phones (D-WS7a); this guards
+    // Desktop-only models are never offered on phones (D-051); this guards
    // any path that could still name one (e.g. a stale selection).
    if (!isModelAllowedOnDevice(targetModel, memoryHint())) return false;
    const device = guessDevice(preference);
@@ -521,7 +521,7 @@ export function App() {
      // use) only where the peak-memory crash risk is real; everywhere else
      // (desktop, low-risk models) the worker loads them in parallel.
      sequentialSessionLoad: isMobileLoadRisk(nextModelId, guessDevice(device)),
-      // D-WS3b: the speech-detection model rides along with the same
+      // D-050: the speech-detection model rides along with the same
      // user-initiated load when "Skip silence" is on.
      vad: longFormRef.current,
    });
@@ -532,7 +532,7 @@ export function App() {
    if (busy || liveStatus !== "idle" || nextModelId === modelId) return;
    setModelId(nextModelId);
    // A "Translate to English" chosen on a model that can translate must not
-    // silently carry over to one that cannot (D-WS4a).
+    // silently carry over to one that cannot (D-044).
    setTask((current) => taskForModel(nextModelId, current));
    setActiveDevice(null);
    setActiveThreads(null);
@@ -714,7 +714,7 @@ export function App() {
        ...(multilingual && { language, task: taskForModel(modelId, task) }),
        prompt: hint,
        // Browser runs only; the worker falls back to the chunked path when
-        // the VAD model isn't available (D-WS3a/D-WS3b).
+        // the VAD model isn't available (D-049/D-050).
        ...(!useServer && { longForm }),
        onPartial: (fragment) => {
          if (fileRun.current === run) setPartialText((text) => text + fragment);
@@ -797,7 +797,7 @@ export function App() {
        language,
        task: taskForModel(modelId, task),
      }),
-      // Same hint for every live pass of the session (D-WS2a).
+      // Same hint for every live pass of the session (D-048).
      ...(hint && { prompt: hint }),
    };

@@ -1027,7 +1027,7 @@ export function App() {
  const serverSelected =
    engineChoice === "server" && auth.user !== null && serverStatus?.available === true;

-  // Phones never see desktop-only models (Parakeet, D-WS7a); the server
+  // Phones never see desktop-only models (Parakeet, D-051); the server
  // engine only offers what the sidecar runs.
  const availableModels = modelsForDevice(
    serverSelected
@@ -1087,7 +1087,7 @@ export function App() {
        serverEngine={serverSelected}
      />

-      {/* Parakeet has no language or task choice (D-WS7a) — say so where
+      {/* Parakeet has no language or task choice (D-051) — say so where
          those controls would be, rather than leaving their absence unexplained. */}
      {parakeetSelected && (
        <p className="model-language-note">
@@ -1109,7 +1109,7 @@ export function App() {
      )}

      {/* Hint and "Skip silence" are Whisper decode options; the Parakeet
-          path (D-WS7a) has neither, so they are hidden rather than ignored. */}
+          path (D-051) has neither, so they are hidden rather than ignored. */}
      {!parakeetSelected && (
        <>
          <VocabularyHint
@@ -1249,7 +1249,7 @@ export function App() {
        }
      />

-      {/* Text translation of the finished transcript (D-WS4b) — any engine,
+      {/* Text translation of the finished transcript (D-045) — any engine,
          file or live; secondary to the transcript itself. */}
      {result && !liveActive && (
        <TranslatePanel
@@ -7,8 +7,8 @@ interface AdvancedOptionsProps {
 /**
 * Technical file-transcription controls, collapsed by default (PRODUCT.md:
 * every control stays available without crowding the primary flow).
- * "Skip silence" turns on the long-form pipeline (D-WS3a), which needs a
- * small speech-detection model downloaded once (D-WS3b).
+ * "Skip silence" turns on the long-form pipeline (D-049), which needs a
+ * small speech-detection model downloaded once (D-050).
 */
 export function AdvancedOptions({
  skipSilence,
@@ -161,7 +161,7 @@ export function ModelPicker({
                    <span className="model-option-name">
                      {model.displayName}
                      {model.family === "parakeet" && (
-                        // Not Whisper (D-WS7a): say what it is good for and
+                        // Not Whisper (D-051): say what it is good for and
                        // what it can't do, in plain words.
                        <span className="model-option-note">
                          Not Whisper · English only · fast on long recordings
@@ -16,7 +16,7 @@ interface StorageManagerProps {
  onChanged: () => void;
 }

-// Whisper weights plus the Opus-MT translation models (D-WS4b).
+// Whisper weights plus the Opus-MT translation models (D-045).
 const MODEL_IDS = [...MODELS.map((model) => model.id), ...TRANSLATION_MODEL_IDS];

 function displayName(modelId: string): string {
@@ -8,7 +8,7 @@ interface TranscriptionOptionsProps {
  onTaskChange: (task: WhisperTask) => void;
  /**
   * Whether the model supports Whisper's own speech → English translation
-   * (`WhisperModel.canTranslate`, D-WS4a). When false the Task control is not
+   * (`WhisperModel.canTranslate`, D-044). When false the Task control is not
   * rendered at all — transcribing is the only thing such a model does.
   */
  canTranslate: boolean;
@@ -18,7 +18,7 @@ interface TranscriptionOptionsProps {
 /**
 * Language + task controls for multilingual models. The parent renders this
 * only when the selected model is multilingual — `.en` models take neither.
- * The task control only appears where the model can translate (D-WS4a).
+ * The task control only appears where the model can translate (D-044).
 */
 export function TranscriptionOptions({
  language,
@@ -65,7 +65,7 @@ interface TranslatePanelProps {
 }

 /**
- * Text translation of a finished transcript with Opus-MT (D-WS4b): a
+ * Text translation of a finished transcript with Opus-MT (D-045): a
 * secondary disclosure under the transcript, separate from Whisper's own
 * "Translate to English" task. Runs on-device in its own worker; the model
 * is downloaded only when the user clicks Translate (D-020).
@@ -10,12 +10,12 @@ interface VocabularyHintProps {
 /**
 * Past this many characters the hint likely exceeds the 120-token prompt
 * budget (a list of names runs ≈ 2.8 characters per token, measured) and gets
- * cut at the end (D-WS2a) — say so rather than silently ignoring the tail.
+ * cut at the end (D-048) — say so rather than silently ignoring the tail.
 */
 const LONG_HINT_CHARS = 300;

 /**
- * Optional "Names and terms" vocabulary hint (D-WS2a). Shown for every model;
+ * Optional "Names and terms" vocabulary hint (D-048). Shown for every model;
 * the parent persists the value (src/lib/hint.ts) and sends it with each
 * transcription as the decoder prompt.
 */
@@ -423,7 +423,7 @@ body {
  flex: 1;
 }

-/* One-line description under a model name (Parakeet, D-WS7a). */
+/* One-line description under a model name (Parakeet, D-051). */
 .model-option-note {
  display: block;
  color: var(--muted);
@@ -526,7 +526,7 @@ body {
  flex-wrap: wrap;
 }

-/* --- VocabularyHint (D-WS2a) --- */
+/* --- VocabularyHint (D-048) --- */

 .vocabulary-hint {
  display: flex;
@@ -580,7 +580,7 @@ body {
  color: var(--muted);
 }

-/* --- AdvancedOptions (long-form "Skip silence", D-WS3a) --- */
+/* --- AdvancedOptions (long-form "Skip silence", D-049) --- */
 .advanced-options summary {
  cursor: pointer;
  color: var(--muted);
@@ -1098,7 +1098,7 @@ code {
 }

 /* --------------------------------------------------------------------------
-   Transcript translation (D-WS4b): a secondary disclosure under the
+   Transcript translation (D-045): a secondary disclosure under the
   transcript, same panel language as .storage-manager.
   -------------------------------------------------------------------------- */

@@ -226,7 +226,7 @@ async function productionServeOptions(): Promise<ServeOptions> {
 * with Bun 1.3: both `bun build` and the dev server leave the call untouched),
 * so worker-like entrypoints are bundled at startup and served from stable
 * URLs. The UI spawns `new Worker("/inference.worker.js")` (and
- * "/translate.worker.js" for transcript translation, D-WS4b); the service
+ * "/translate.worker.js" for transcript translation, D-045); the service
 * worker registers as "/sw.js".
 */
 async function bundleEntrypoint(relativePath: string): Promise<string> {
@@ -237,7 +237,7 @@ async function bundleEntrypoint(relativePath: string): Promise<string> {
    // Never minify: sw.ts relies on its placeholder string literals surviving
    // bundling verbatim (src/pwa/precache.ts), and dev builds favor debuggability.
    minify: false,
-    // One onnxruntime in the worker bundle (D-WS7a); harmless for sw.ts.
+    // One onnxruntime in the worker bundle (D-051); harmless for sw.ts.
    plugins: [ortAliasPlugin],
  });
  const artifact = build.outputs[0];
@@ -254,7 +254,7 @@ async function developmentServeOptions(): Promise<ServeOptions> {
  const { default: index } = await import("./index.html");

  const workerScript = await bundleEntrypoint("./worker/inference.worker.ts");
-  // Transcript translation runs in its own worker (D-WS4b).
+  // Transcript translation runs in its own worker (D-045).
  const translateWorkerScript = await bundleEntrypoint("./worker/translate.worker.ts");
  // Dev SW: empty precache list — assets are unhashed here, so the shell is
  // runtime-cached network-first instead (see src/pwa/sw.ts).
@@ -38,7 +38,7 @@ export interface EngineTranscribeOptions {
  /** transcribe | translate (multilingual models only). */
  task?: WhisperTask;
  /**
-   * "Names and terms" vocabulary hint, raw user text (D-WS2a). Applies to
+   * "Names and terms" vocabulary hint, raw user text (D-048). Applies to
   * every Whisper model, `.en` ones included; empty/undefined = no prompt.
   */
  prompt?: string;
@@ -61,7 +61,7 @@ export interface EngineTranscribeOptions {
  signal?: AbortSignal;
  /**
   * Browser engine only: use the long-form pipeline when the worker has the
-   * VAD model (D-WS3a). The server engine runs faster-whisper's own pipeline.
+   * VAD model (D-049). The server engine runs faster-whisper's own pipeline.
   */
  longForm?: boolean;
 }
@@ -294,7 +294,7 @@ export class ServerEngine implements Engine {
      params.set("language", options.language);
    }
    if (options.task !== undefined) params.set("task", options.task);
-    // Validated + normalized server-side (parseLanguageAndTask, D-WS2a).
+    // Validated + normalized server-side (parseLanguageAndTask, D-048).
    if (options.prompt) params.set("prompt", options.prompt);

    let response: Response;
@@ -17,7 +17,7 @@
 */

 /**
- * `"translate"` covers text translation of a finished transcript (D-WS4b):
+ * `"translate"` covers text translation of a finished transcript (D-045):
 * both its one-time model download and the translation pass, which run in
 * their own worker — only the suggestions differ from the Whisper stages.
 */
@@ -1,5 +1,5 @@
 /**
- * The optional "Names and terms" vocabulary hint (D-WS2a): free text the user
+ * The optional "Names and terms" vocabulary hint (D-048): free text the user
 * types once, remembered in this browser, and fed to Whisper as a decoder
 * prompt (`<|startofprev|> hint …`) so names and jargon come out spelled the
 * way the user spells them.
@@ -27,7 +27,7 @@ export const HINT_MAX_CHARS = 1000;
 * faster-whisper cap), but measured on en_alien a prompt that long garbles
 * small models: base.en WER 0.339 → 0.476 with a realistic 218-token list and
 * → 1.05 with a synthetic one, while ≤ 120 tokens stays at or below the
- * no-hint WER (D-WS2a).
+ * no-hint WER (D-048).
 */
 export const PROMPT_TOKEN_BUDGET = 120;

@@ -74,12 +74,12 @@ function defaultStorage(): StorageLike | null {

 /**
 * Token ids for the hint as a Whisper prompt (without `<|startofprev|>`),
- * truncated predictably to `budget` tokens (D-WS2a):
+ * truncated predictably to `budget` tokens (D-048):
 *
 * 1. Prompt text = `" " + normalizeHint(hint)` (Whisper prompts start with a
 *    space), ending in a full stop unless it already ends in . ? or ! —
 *    measured: an unterminated list makes Whisper drop punctuation and, on
- *    smaller models, fall into repetition loops (D-WS2a).
+ *    smaller models, fall into repetition loops (D-048).
 * 2. Over budget → keep the longest prefix of whole space-separated words that
 *    fits (full stop included), a dangling "," / ";" removed: the start of the
 *    list is kept, the end dropped, never mid-word.
@@ -1,6 +1,6 @@
 /**
 * The "Skip silence" file-transcription setting (long-form pipeline,
- * D-WS3a): remembered in localStorage like the engine choice.
+ * D-049): remembered in localStorage like the engine choice.
 */

 import type { StorageLike } from "./engine";
@@ -10,7 +10,7 @@ export const LONG_FORM_KEY = "whisper-long-form";
 /**
 * Off by default: the spec's rollout rule makes the long-form pipeline the
 * default only if it is faster and not less accurate on every measurement
- * clip — see D-WS3a for the numbers that decided this.
+ * clip — see D-049 for the numbers that decided this.
 */
 export const LONG_FORM_DEFAULT = false;

@@ -90,7 +90,7 @@ export function shouldWarnBeforeLoad(hint: MobileMemoryHint, peakBytes: number):
 }

 /**
- * Desktop-only models (D-WS7a): Parakeet's ~1.3 GB load peak (2 × the
+ * Desktop-only models (D-051): Parakeet's ~1.3 GB load peak (2 × the
 * 652 MB int8 encoder, D-027's arithmetic) exceeds every phone budget the
 * D-029 table considers safe, even at deviceMemory 8 (800 MB). Unlike the
 * Whisper models — where a confirm lets the user judge their device — it is
@@ -36,14 +36,14 @@ export interface DtypeDefault {
 }

 /**
- * Which inference path a catalog entry runs on (D-WS7a). Whisper models use
+ * Which inference path a catalog entry runs on (D-051). Whisper models use
 * the Transformers.js ASR pipeline; Parakeet uses the worker's own TDT path
 * (src/worker/parakeetEngine.ts). Absent ⇒ "whisper", so the Whisper entries
 * carry no field at all.
 */
 export type ModelFamily = "whisper" | "parakeet";

-/** Exact weight-file sizes per dtype, for entries whose files we know (D-WS7a). */
+/** Exact weight-file sizes per dtype, for entries whose files we know (D-051). */
 export interface ExactModelBytes {
  encoder: Partial<Record<WhisperDtype, number>>;
  decoder: Partial<Record<WhisperDtype, number>>;
@@ -54,7 +54,7 @@ export interface ExactModelBytes {
 export interface WhisperModel {
  /** Hugging Face Hub id, e.g. "onnx-community/whisper-tiny.en". */
  id: string;
-  /** Inference path (D-WS7a); absent ⇒ "whisper". */
+  /** Inference path (D-051); absent ⇒ "whisper". */
  family?: ModelFamily;
  /**
   * Real file sizes; when present they replace the parameter-count estimate
@@ -67,11 +67,11 @@ export interface WhisperModel {
  multilingual: boolean;
  /**
   * Whether Whisper's own `task: "translate"` (speech → English text) is
-   * offered (D-WS4a). False for every `.en` model (no task token) and for
+   * offered (D-044). False for every `.en` model (no task token) and for
   * large-v3-turbo, which OpenAI did not train for translation — and for any
   * turbo-based fine-tune not tested for it. The UI hides "Translate to
   * English" where false; text translation of a finished transcript
-   * (D-WS4b) works for every model regardless.
+   * (D-045) works for every model regardless.
   */
  canTranslate: boolean;
  /** Encoder parameter count, in millions (drives size estimates). */
@@ -86,7 +86,7 @@ export interface WhisperModel {
   * CTranslate2 conversions on the Hub (`Systran/faster-whisper-*`;
   * large-v3-turbo → `mobiuslabsgmbh/faster-whisper-large-v3-turbo`), so the
   * pinned library version stays the single authority on the exact repo.
-   * Absent for browser-only families (Parakeet, D-WS7a): the server engine
+   * Absent for browser-only families (Parakeet, D-051): the server engine
   * answers 400 for them and the picker hides them while it is selected.
   */
  fasterWhisperModel?: string;
@@ -212,12 +212,12 @@ const ACCEL_Q4: DtypeDefault = {
 };

 // ---------------------------------------------------------------------------
-// Danish fine-tunes (D-WS6a)
+// Danish fine-tunes (D-046)
 // ---------------------------------------------------------------------------

 /**
 * Hugging Face account that hosts our conversions of the Danish Whisper
- * fine-tunes (D-WS6a). Both the ONNX repo (browser engine) and the
+ * fine-tunes (D-046). Both the ONNX repo (browser engine) and the
 * CTranslate2 int8 repo (server engine) hang off this one value. Only
 * whisper-small-danish is published: the turbo fine-tunes (turbo-danish,
 * edda-v0.2) showed no clear gain over stock turbo on our Danish clip.
@@ -225,14 +225,14 @@ const ACCEL_Q4: DtypeDefault = {
 export const DANISH_MODELS_HF_OWNER = "lucasw89";

 /**
- * WASM default for the Danish conversions (D-WS6a): measured on our own files
+ * WASM default for the Danish conversions (D-046): measured on our own files
 * (da_festuge, onnxruntime-node CPU), q4 matched or beat q8 on WER and was
 * faster for all three — small included, unlike the stock small (q8).
 */
 const DANISH_WASM_Q4: DtypeDefault = {
  dtype: "q4",
  verified: false,
-  note: "q4: best measured WER and fastest of q8/q4 on our Danish conversions (D-WS6a).",
+  note: "q4: best measured WER and fastest of q8/q4 on our Danish conversions (D-046).",
 };

 /** Catalog id + matching faster-whisper repo for one published conversion. */
@@ -247,7 +247,7 @@ function danishRepos(name: string): Pick<WhisperModel, "id" | "fasterWhisperMode
 // The catalog (D-006: curated, not exhaustive)
 // ---------------------------------------------------------------------------

-/** Parakeet TDT 0.6B v3 ONNX repo and the revision its file sizes were taken from (D-WS7a). */
+/** Parakeet TDT 0.6B v3 ONNX repo and the revision its file sizes were taken from (D-051). */
 export const PARAKEET_MODEL_ID = "ysdede/parakeet-tdt-0.6b-v3-onnx";
 export const PARAKEET_REVISION = "e5356bf6e24bdb7a269536b0e9e7869bc034b893";

@@ -347,13 +347,13 @@ export const MODELS: readonly WhisperModel[] = [
    displayName: "Whisper Large v3 Turbo",
    multilingual: true,
    // Not trained for translation (OpenAI's turbo model card); its
-    // "translate" output is unreliable (D-WS4a).
+    // "translate" output is unreliable (D-044).
    canTranslate: false,
    encoderParamsM: 635,
    decoderParamsM: 174,
    defaults: { wasm: WASM_Q4, webgpu: WEBGPU_FP16_Q4, webnn: WEBNN_FP16 },
  },
-  // Danish fine-tune (D-WS6a): unchanged Whisper architecture, so the
+  // Danish fine-tune (D-046): unchanged Whisper architecture, so the
  // parameter counts are its base model's. Conversions by us, weights by
  // the original authors (credited in each published repo's model card).
  {
@@ -361,7 +361,7 @@ export const MODELS: readonly WhisperModel[] = [
    ...danishRepos("whisper-small-danish"),
    displayName: "Whisper Small (Danish)",
    multilingual: true,
-    // Translation untested on these fine-tunes (spec 001, D-WS4a).
+    // Translation untested on these fine-tunes (spec 001, D-044).
    canTranslate: false,
    encoderParamsM: 88,
    decoderParamsM: 156,
@@ -369,7 +369,7 @@ export const MODELS: readonly WhisperModel[] = [
  },
  {
    // NVIDIA Parakeet TDT 0.6B v3 (CC-BY-4.0), NeMo→ONNX conversion by
-    // ysdede, run through parakeet.js on our onnxruntime-web (D-WS7a).
+    // ysdede, run through parakeet.js on our onnxruntime-web (D-051).
    // English-only in this app (the model itself knows 25 European
    // languages, but only English is measured here). Desktop-only: its
    // ~1.3 GB load peak is far past every phone budget (D-027–D-029), so
@@ -391,10 +391,10 @@ export const MODELS: readonly WhisperModel[] = [
      wasm: {
        dtype: "q8",
        verified: false,
-        note: "int8 encoder + decoder: the measured WASM path (D-WS7a).",
+        note: "int8 encoder + decoder: the measured WASM path (D-051).",
      },
      // Only used when WebGPU is chosen explicitly — "auto" never picks it
-      // for Parakeet (D-WS7a): int8 on WebGPU is reported broken, and fp16
+      // for Parakeet (D-051): int8 on WebGPU is reported broken, and fp16
      // is a separate 1.24 GB download. The decoder stays int8 on WASM.
      webgpu: {
        dtype: { encoder_model: "fp16", decoder_model_merged: "q8" },
@@ -424,7 +424,7 @@ export function getFasterWhisperModel(id: string): string | undefined {
 }

 /**
- * The task a run on `modelId` should actually use (D-WS4a): a "translate"
+ * The task a run on `modelId` should actually use (D-044): a "translate"
 * left over from a model that could translate falls back to "transcribe" on
 * one that cannot (large-v3-turbo, `.en`, unknown ids). App applies it on
 * every model switch and again when building each run's options.
@@ -1,5 +1,5 @@
 /**
- * Pure logic of the Parakeet TDT engine path (Decisions.md D-WS7a): window
+ * Pure logic of the Parakeet TDT engine path (Decisions.md D-051): window
 * planning, word → transcript-chunk mapping, the vocab tokenizer, and the
 * pinned asset URLs. No runtime, no browser globals — the worker side
 * (src/worker/parakeetEngine.ts) injects the actual recognizer.
@@ -1,5 +1,5 @@
 /**
- * Text translation of a finished transcript (D-WS4b) — the pure half.
+ * Text translation of a finished transcript (D-045) — the pure half.
 *
 * Opus-MT (MarianMT) runs in `src/worker/translate.worker.ts`; everything
 * here is planning that needs no model: which checkpoint serves a language
@@ -237,7 +237,7 @@ export interface TranslateUnitsDeps extends SplitOptions {
 * One piece per model call. Batching would be ~4x faster, but Transformers.js
 * 4.2's batched Marian generation is broken: every row of a padded batch
 * comes back as row 0's translation (measured 2026-10-29 on
- * onnxruntime-node, q8 and fp32, any graph optimisation level — D-WS4b).
+ * onnxruntime-node, q8 and fp32, any graph optimisation level — D-045).
 * Re-measure with `scripts/bench-translate.ts` after a Transformers.js
 * upgrade before raising this.
 */
@@ -25,7 +25,7 @@ export interface LiveSessionData {
  model: string;
  language?: string;
  task: WhisperTask;
-  /** Normalized vocabulary hint, applied to every pass (D-WS2a). */
+  /** Normalized vocabulary hint, applied to every pass (D-048). */
  prompt?: string;
  tempDir: string;
 }
@@ -62,7 +62,7 @@ export function createLiveWebSocket(deps: LiveWebSocketDeps) {
                audioPath,
                language: ws.data.language,
                task: ws.data.task,
-                // Short passes: sequential decode, no batching/VAD (D-WS5a).
+                // Short passes: sequential decode, no batching/VAD (D-047).
                mode: "live",
                prompt: ws.data.prompt,
              },
@@ -50,11 +50,11 @@ export interface SidecarTranscribeRequest {
  task?: WhisperTask;
  /**
   * "file" (default): whole uploads — batched inference over VAD speech
-   * chunks. "live": short server-live passes — sequential, no VAD (D-WS5a).
+   * chunks. "live": short server-live passes — sequential, no VAD (D-047).
   */
  mode?: SidecarMode;
  /**
-   * Normalized "Names and terms" hint (D-WS2a); the sidecar truncates it to
+   * Normalized "Names and terms" hint (D-048); the sidecar truncates it to
   * Whisper's prompt budget with faster-whisper's own tokenizer.
   */
  prompt?: string;
@@ -349,7 +349,7 @@ interface TranscribeParams {
 * unvalidated string into the sidecar protocol just because it skips
 * `parseParams` below (D-032). A string return is the 400 error text.
 *
- * `prompt` is the raw "Names and terms" hint (D-WS2a): at most
+ * `prompt` is the raw "Names and terms" hint (D-048): at most
 * HINT_MAX_CHARS (the UI's textarea limit), normalized exactly as the browser
 * worker does; empty after normalization = no prompt.
 */
@@ -18,7 +18,7 @@ import {
 } from "@huggingface/transformers";
 // The very onnxruntime-web module Transformers.js imports (its web build
 // imports "onnxruntime-web/webgpu"), so the Parakeet path shares one runtime,
-// one wasm binary from /ort/, and one env (D-WS7a).
+// one wasm binary from /ort/, and one env (D-051).
 import * as ort from "onnxruntime-web/webgpu";
 import { getModel, modelFamily, resolveDefaultDtype } from "../lib/models";
 import { PARAKEET_VOCAB_FILE, parakeetFileName, parakeetFileUrl } from "../lib/parakeet";
@@ -222,7 +222,7 @@ async function probeCapabilities(
 // Pipeline lifecycle (load, switch, dispose)
 // ---------------------------------------------------------------------------

-/** A loaded model of either family (D-WS7a): the worker dispatches on it. */
+/** A loaded model of either family (D-051): the worker dispatches on it. */
 type LoadedModel =
  | { family: "whisper"; asr: AutomaticSpeechRecognitionPipeline }
  | { family: "parakeet"; engine: ParakeetEngine };
@@ -297,7 +297,7 @@ async function createPipeline(
        // graph optimization level. D-014's WASM "basic" cap worked around a
        // level-2 QDQ bug in ort-web 1.26.0-dev that rejected the 8-bit
        // decoders; the 1.31.0-dev build bundled with Transformers.js 4.3.2
-        // loads them at every level (D-WS1c, pinned in
+        // loads them at every level (D-043, pinned in
        // tests/transformers-runtime.test.ts).
        progress_callback: forwardProgress,
        // D-027/D-029: sequential sub-model session loading trades load speed
@@ -324,7 +324,7 @@ async function createPipeline(
 }

 /**
- * Parakeet load (D-WS7a): fetch vocab + decoder + encoder through the
+ * Parakeet load (D-051): fetch vocab + decoder + encoder through the
 * Cache-API downloader (same cache and progress messages as Transformers.js,
 * D-005), then create the sessions on our shared onnxruntime-web. Device
 * policy is `parakeetDeviceAttempts` (WASM unless WebGPU was chosen
@@ -412,7 +412,7 @@ function load(message: LoadMessage): Promise<LoadedModel> {
 }

 /**
- * Silero VAD for long-form file passes (D-WS3a/D-WS3b). Loaded on request
+ * Silero VAD for long-form file passes (D-049/D-050). Loaded on request
 * (`load.vad`) once the Whisper pipeline is up — after it, so the Whisper
 * attempt decides onnxruntime's thread pool (D-023/D-025). Offline it is
 * used only if already cached: no network request (Review Focus 3). A failed
@@ -452,7 +452,7 @@ const CHUNK_LENGTH_S = 30;
 const STRIDE_LENGTH_S = 5;

 /**
- * Parakeet pass (D-WS7a). English-only: `language`/`task` are ignored (the UI
+ * Parakeet pass (D-051). English-only: `language`/`task` are ignored (the UI
 * never sends them for a non-multilingual model). File-flow passes stream one
 * `partial` + one `transcribe-progress` per ≤ 30 s window — absolute seconds
 * straight from the window plan, so no TranscribeProgressTracker is needed.
@@ -513,7 +513,7 @@ async function transcribe(message: TranscribeMessage): Promise<void> {
    return_timestamps: true,
    // With return_timestamps and no explicit max_new_tokens, Transformers.js
    // 4.2–4.3 Whisper generate() takes a seek-loop path (D-021; re-measured
-    // on 4.3.2, D-WS1b) that re-decodes each window's tail (1–3 generate
+    // on 4.3.2, D-042) that re-decodes each window's tail (1–3 generate
    // passes per window). Under this pipeline's own 30 s/5 s chunking that
    // is pure overhead: ~1.4× slower on long
    // files, duplicated boundary text, on_finalize firing once per pass
@@ -575,7 +575,7 @@ async function transcribe(message: TranscribeMessage): Promise<void> {
  if (message.task !== undefined) {
    options.task = message.task;
  }
-  // "Names and terms" hint (D-WS2a): one prefix per pass, applied by the
+  // "Names and terms" hint (D-048): one prefix per pass, applied by the
  // generate() wrapper to every 30 s window (Transformers.js ignores
  // `prompt_ids`, see src/worker/prompt.ts). The wrapper is idempotent.
  const promptPrefix = buildPromptPrefix(
@@ -603,7 +603,7 @@ async function transcribe(message: TranscribeMessage): Promise<void> {
 }

 /**
- * A file pass through the long-form pipeline (D-WS3a): same partial /
+ * A file pass through the long-form pipeline (D-049): same partial /
 * transcribe-progress / result messages as the chunked path.
 */
 async function transcribeLongFormPass(
@@ -625,7 +625,7 @@ async function transcribeLongFormPass(
      ...LONG_FORM_DEFAULTS,
      ...(multilingual && message.language !== undefined && { language: message.language }),
      ...(multilingual && message.task !== undefined && { task: message.task }),
-      // Names-and-terms hint on every chunk, like the chunked path (D-WS2a).
+      // Names-and-terms hint on every chunk, like the chunked path (D-048).
      ...(message.prompt && { prompt: message.prompt }),
      onPartial: (text) => post({ type: "partial", text, requestId: message.requestId }),
      onProgress: (processedSeconds, totalSeconds) =>
@@ -1,4 +1,4 @@
-/** Long-form file transcription pipeline (D-WS3a). See pipeline.ts. */
+/** Long-form file transcription pipeline (D-049). See pipeline.ts. */
 export { DEFAULT_GUARD, LONG_FORM_DEFAULTS, transcribeLongForm } from "./pipeline";
 export { DEFAULT_FALLBACK, type FallbackOptions } from "./quality";
 export type {
@@ -1,5 +1,5 @@
 /**
- * Long-form file transcription (D-WS3a), mirroring faster-whisper's batched
+ * Long-form file transcription (D-049), mirroring faster-whisper's batched
 * pipeline without the batching: Silero VAD drops silence and cuts the audio
 * on speech boundaries into chunks ≤ 30 s; each chunk is one Whisper decode
 * (no overlapping windows to stitch); the language is detected once on the
@@ -34,10 +34,10 @@ import {
 /** Whisper's window: no chunk is longer. */
 const MAX_CHUNK_SECONDS = 30;

-/** No guard: Whisper's own decoding (the measured default, D-WS3a). */
+/** No guard: Whisper's own decoding (the measured default, D-049). */
 export const DEFAULT_GUARD: RepetitionGuard = { noRepeatNgramSize: 0, repetitionPenalty: 1 };

-/** The tuning the worker runs with — chosen by measurement (D-WS3a). */
+/** The tuning the worker runs with — chosen by measurement (D-049). */
 export const LONG_FORM_DEFAULTS: Pick<LongFormOptions, "guard" | "fallback" | "vad"> = {
  guard: DEFAULT_GUARD,
  fallback: false,
@@ -1,6 +1,6 @@
 /**
 * Per-chunk decode quality checks for the long-form pipeline's temperature
- * fallback (D-WS3a) — faster-whisper's `generate_with_fallback` criteria.
+ * fallback (D-049) — faster-whisper's `generate_with_fallback` criteria.
 */

 export interface FallbackOptions {
@@ -1,5 +1,5 @@
 /**
- * `LongFormDeps` over the existing Transformers.js runtime (D-WS3a/D-WS3b):
+ * `LongFormDeps` over the existing Transformers.js runtime (D-049/D-050):
 * Silero VAD runs as a plain ONNX model on the same onnxruntime instance the
 * Whisper pipeline uses, and Whisper decodes go through the already-loaded
 * ASR pipeline — no second inference runtime. Free of worker globals so
@@ -30,7 +30,7 @@ import { SAMPLE_RATE, VAD_WINDOW_SAMPLES } from "./vad";
 /**
 * Silero VAD v5, ONNX export by onnx-community (MIT; `onnx/model.onnx`,
 * fp32, 2.2 MB). Fetched from the Hub like Whisper weights — CORS-enabled,
- * so COEP allows it — and kept in the same Cache API store (D-005, D-WS3b).
+ * so COEP allows it — and kept in the same Cache API store (D-005, D-050).
 */
 export const VAD_MODEL_ID = "onnx-community/silero-vad";

@@ -60,7 +60,7 @@ async function vadInBrowserCache(): Promise<boolean> {

 export async function loadSileroVad(options: LoadSileroVadOptions): Promise<SileroVad> {
  if (options.localOnly && !(await (options.isCached ?? vadInBrowserCache)())) {
-    throw new Error("The speech-detection model is not downloaded (D-WS3b).");
+    throw new Error("The speech-detection model is not downloaded (D-050).");
  }
  const model = await AutoModel.from_pretrained(VAD_MODEL_ID, {
    // The repo ships no config.json: the model is a bare ONNX graph.
@@ -200,7 +200,7 @@ export function createWhisperDeps(
      if (request.task !== undefined) generate.task = request.task;
    }
    if (request.prompt) {
-      // Same decoder-prompt mechanism as the chunked path (D-WS2a,
+      // Same decoder-prompt mechanism as the chunked path (D-048,
      // src/worker/prompt.ts); the wrapper strips the prompt from the output.
      const prefix = promptPrefixFor(request.prompt);
      if (prefix.length > 0) generate[PROMPT_PREFIX_KWARG] = prefix;
@@ -1,5 +1,5 @@
 /**
- * The long-form pipeline's interface (D-WS3a, coordination plan 001-0). The
+ * The long-form pipeline's interface (D-049, coordination plan 001-0). The
 * pipeline is model-agnostic: everything that runs a model comes in through
 * `LongFormDeps`, so a different implementation (a future Rust→WASM build,
 * or batched decoding) can replace either side without touching the other.
@@ -46,7 +46,7 @@ export interface LongFormDeps {
  detectLanguage?(audio: Float32Array): Promise<string>;
 }

-/** Decoder-side repetition guard (D-035's proposal; D-WS3a measured it). */
+/** Decoder-side repetition guard (D-035's proposal; D-049 measured it). */
 export interface RepetitionGuard {
  noRepeatNgramSize: number;
  repetitionPenalty: number;
@@ -1,5 +1,5 @@
 /**
- * Speech segmentation for the long-form pipeline (D-WS3a). Pure: turns
+ * Speech segmentation for the long-form pipeline (D-049). Pure: turns
 * per-window Silero VAD speech probabilities into speech spans, and packs the
 * spans into ≤ 30 s chunks cut on speech boundaries. A port of faster-whisper
 * 1.2.1 `vad.py` (`get_speech_timestamps` after the model call, and the
@@ -1,6 +1,6 @@
 /**
 * Weight-file download + cache for engine paths that don't go through
- * Transformers.js's own loader (Parakeet, D-WS7a).
+ * Transformers.js's own loader (Parakeet, D-051).
 *
 * Files land in the same Cache Storage cache Transformers.js uses
 * (`transformers-cache`, D-005), keyed by their Hub URL, so the downloaded
@@ -1,6 +1,6 @@
 /**
 * Opus-MT (MarianMT) through Transformers.js — the model half of transcript
- * translation (D-WS4b). Shared by the browser translate worker and the
+ * translation (D-045). Shared by the browser translate worker and the
 * onnxruntime-node measurement script / real-model test, so all three run
 * the exact same pipeline options.
 */
@@ -1,6 +1,6 @@
 /**
 * Stand-in for the bare `onnxruntime-web` import inside parakeet.js
- * (D-WS7a). parakeet.js depends on its own onnxruntime-web 1.24.1, imported
+ * (D-051). parakeet.js depends on its own onnxruntime-web 1.24.1, imported
 * only by its `initOrt` helper, which this app never calls — but a bundler
 * would still pull that second runtime into the worker. `scripts/ortAlias.ts`
 * redirects the import here, to the one runtime Transformers.js uses.
@@ -1,5 +1,5 @@
 /**
- * Parakeet TDT engine (Decisions.md D-WS7a): the non-Whisper inference path
+ * Parakeet TDT engine (Decisions.md D-051): the non-Whisper inference path
 * of the browser engine, used by the inference worker for catalog entries
 * with `family: "parakeet"`.
 *
@@ -1,5 +1,5 @@
 /**
- * Whisper decoder prompts (the "Names and terms" hint, D-WS2a) on top of the
+ * Whisper decoder prompts (the "Names and terms" hint, D-048) on top of the
 * Transformers.js ASR pipeline.
 *
 * Transformers.js 4.2.0 (and 4.3.2) declare `prompt_ids` in the Whisper
@@ -59,7 +59,7 @@ export interface LoadMessage {
  sequentialSessionLoad?: boolean;
  /**
   * Also make the Silero VAD model (2.2 MB) available for long-form file
-   * transcription (D-WS3a/D-WS3b): cache-first, fetched from the Hub only
+   * transcription (D-049/D-050): cache-first, fetched from the Hub only
   * when online — offline it is used only if already cached. Sent with the
   * same user-initiated loads as the Whisper weights (D-020).
   */
@@ -96,14 +96,14 @@ export interface TranscribeMessage {
  stream?: boolean;
  /**
   * Optional "Names and terms" vocabulary hint, raw as the user typed it
-   * (D-WS2a). The worker normalizes it, truncates it to Whisper's prompt
+   * (D-048). The worker normalizes it, truncates it to Whisper's prompt
   * budget and feeds it as a decoder prompt to every window of this pass
   * (src/worker/prompt.ts). Empty/omitted = no prompt.
   */
  prompt?: string;
  /**
   * Run this file pass through the long-form pipeline (VAD-cut chunks,
-   * language pinned once, D-WS3a). Honoured only when the VAD model loaded;
+   * language pinned once, D-049). Honoured only when the VAD model loaded;
   * otherwise the classic 30 s/5 s chunked path runs (`choosePipeline`).
   */
  longForm?: boolean;
@@ -288,7 +288,7 @@ export function deviceAttempts(
 }

 /**
- * Which file-transcription path a pass takes (D-WS3a/D-WS3b): long-form only
+ * Which file-transcription path a pass takes (D-049/D-050): long-form only
 * when the pass asked for it and the VAD model is loaded. A missing VAD model
 * (offline, never downloaded) silently keeps the classic chunked path — it is
 * a fallback, never an error.
@@ -301,7 +301,7 @@ export function choosePipeline(input: {
 }

 /**
- * Device attempts for the Parakeet engine path (D-WS7a). Narrower than
+ * Device attempts for the Parakeet engine path (D-051). Narrower than
 * `deviceAttempts`: "auto" always means WASM int8 — the only measured path —
 * because WebGPU int8 is reported broken and WebGPU needs the separate
 * 1.24 GB fp16 encoder, so auto-selecting it would silently double the
@@ -496,7 +496,7 @@ export function applyLoadProgress(

 /**
 * One generate pass's slice of the audio, for
- * `TranscribeProgressTracker.forWindows` (variable-length chunks, D-WS3a).
+ * `TranscribeProgressTracker.forWindows` (variable-length chunks, D-049).
 */
 export interface ProgressWindow {
  /** Window-relative seconds → absolute seconds on the file's timeline. */
@@ -521,7 +521,7 @@ export interface ProgressWindow {
 * ends. This tracker turns those into absolute seconds that are monotone
 * (windows overlap, so a later window's early timestamps can precede an
 * earlier window's late ones — regressions are swallowed) and clamped to
- * [0, totalSeconds]. The long-form pipeline (D-WS3a) uses `forWindows`
+ * [0, totalSeconds]. The long-form pipeline (D-049) uses `forWindows`
 * instead: its windows are variable-length speech chunks with skipped
 * silence between them.
 */
@@ -551,7 +551,7 @@ export class TranscribeProgressTracker {

  /**
   * Variable-length windows — the long-form pipeline's speech chunks
-   * (D-WS3a): each maps its own relative time onto the file's timeline, and
+   * (D-049): each maps its own relative time onto the file's timeline, and
   * skipped silence between chunks counts as processed once a chunk ends.
   */
  static forWindows(
@@ -1,5 +1,5 @@
 /**
- * Transcript-translation Web Worker (D-WS4b): Opus-MT through
+ * Transcript-translation Web Worker (D-045): Opus-MT through
 * Transformers.js, off the main thread (D-003).
 *
 * A dedicated worker rather than a message type on the Whisper worker:
@@ -1,6 +1,6 @@
 /**
 * Typed postMessage protocol between the UI and the transcript-translation
- * worker (`src/worker/translate.worker.ts`, D-WS4b). Kept apart from the
+ * worker (`src/worker/translate.worker.ts`, D-045). Kept apart from the
 * Whisper worker's protocol: separate worker, separate lifetime. Free of
 * browser globals so `bun test` covers the guards.
 */
@@ -77,7 +77,7 @@ describe("BrowserEngine", () => {
    expect("modelId" in message).toBe(false);
  });

-  test("forwards the vocabulary hint as `prompt`, and omits an empty one (D-WS2a)", () => {
+  test("forwards the vocabulary hint as `prompt`, and omits an empty one (D-048)", () => {
    const worker = new FakeWorker();
    const engine = new BrowserEngine(worker);
    void engine.transcribe(new Float32Array(4), { prompt: "Maxwell\nVoid" });
@@ -88,7 +88,7 @@ describe("BrowserEngine", () => {
    expect("prompt" in worker.posted[2]!.message).toBe(false);
  });

-  test("forwards the long-form request only when set (D-WS3a)", () => {
+  test("forwards the long-form request only when set (D-049)", () => {
    const worker = new FakeWorker();
    const engine = new BrowserEngine(worker);
    void engine.transcribe(new Float32Array(4), { longForm: true });
@@ -99,7 +99,7 @@ describe("classifyError", () => {
  });
 });

-describe("classifyError — transcript translation (D-WS4b)", () => {
+describe("classifyError — transcript translation (D-045)", () => {
  test("out of memory while translating suggests closing tabs, not a smaller Whisper model", () => {
    const classified = classifyError("RangeError: Out of memory", { stage: "translate" });
    expect(classified.kind).toBe("out-of-memory");
@@ -1,7 +1,7 @@
 """faster_whisper stand-in that reports how server-py/sidecar.py called it.

 Lets tests/transcribe.test.ts check the real sidecar's request routing
-(D-WS5a: file → BatchedInferencePipeline + VAD, live → sequential) without
+(D-047: file → BatchedInferencePipeline + VAD, live → sequential) without
 CTranslate2 or model weights. Every segment's text describes the call:

  " seq model=<name> loads=<n> <sorted kwargs>"      WhisperModel.transcribe
@@ -2,7 +2,7 @@ import { describe, expect, test } from "bun:test";
 import { join } from "node:path";
 import { fitHintToBudget, normalizeHint, PROMPT_TOKEN_BUDGET } from "../src/lib/hint";

-// D-WS2a: the server engine truncates the hint in Python (server-py/sidecar.py,
+// D-048: the server engine truncates the hint in Python (server-py/sidecar.py,
 // with faster-whisper's tokenizer) and the browser engine in TypeScript
 // (src/lib/hint.ts, with the Transformers.js tokenizer). Both must apply the
 // same rule; with the same deterministic fake tokenizer they must agree
@@ -67,13 +67,13 @@ describe("hint storage", () => {
  });
 });

-describe("fitHintToBudget (D-WS2a truncation rule)", () => {
+describe("fitHintToBudget (D-048 truncation rule)", () => {
  test("empty hint means no prompt", () => {
    expect(fitHintToBudget("  \n", charEncode)).toEqual([]);
  });

  test("a hint within budget is encoded as a space-led sentence ending in a full stop", () => {
-    // Measured (D-WS2a): an unterminated list prompt suppresses punctuation
+    // Measured (D-048): an unterminated list prompt suppresses punctuation
    // and triggers repetition loops on smaller models; a full stop does not.
    expect(fitHintToBudget("Maxwell\nVoices of the Void", charEncode)).toEqual(
      charEncode(" Maxwell, Voices of the Void."),
@@ -329,7 +329,7 @@ describe("live pipeline accuracy (real model, real recordings)", () => {
 });

 // ---------------------------------------------------------------------------
-// Vocabulary hint (D-WS2a): the worker's real prompt code on real weights
+// Vocabulary hint (D-048): the worker's real prompt code on real weights
 // ---------------------------------------------------------------------------

 const HINT = "Maxwell, Voices of the Void, Slenderman, RuneScape, UFO";
@@ -372,7 +372,7 @@ async function transcribeWithHint(
  return single.text ?? "";
 }

-describe("vocabulary hint (real model, D-WS2a)", () => {
+describe("vocabulary hint (real model, D-048)", () => {
  test.skipIf(!available)(
    "a hint keeps the transcript intact, keeps the hinted names, and never leaks into the text",
    async () => {
@@ -429,7 +429,7 @@ describe("vocabulary hint (real model, D-WS2a)", () => {
 });

 // ---------------------------------------------------------------------------
-// File-mode accuracy of the long-form pipeline (D-WS3a): Silero VAD chunks,
+// File-mode accuracy of the long-form pipeline (D-049): Silero VAD chunks,
 // one decode per ≤ 30 s chunk — the path behind "Skip silence". Same
 // per-clip bars as the live and chunked file paths above.
 // ---------------------------------------------------------------------------
@@ -462,7 +462,7 @@ describe("long-form file pipeline accuracy (real model, real recordings)", () =>
          if (clip.language !== "en") {
            // "Auto-detect" in the UI: the chunked path forces English here
            // (Transformers.js 4.2 has no detection); long-form detects once
-            // on the first speech chunk and pins it (D-WS3a).
+            // on the first speech chunk and pins it (D-049).
            // Detection must not cost an extra encoder pass: the first
            // chunk's decode reuses the encoder output detection computed.
            const model = asr.model as unknown as Record<string, unknown>;
@@ -18,7 +18,7 @@ function memoryStorage(initial: Record<string, string> = {}) {
  };
 }

-describe("long-form setting persistence (D-WS3a)", () => {
+describe("long-form setting persistence (D-049)", () => {
  test("defaults when nothing is stored or storage is unavailable", () => {
    expect(loadLongFormChoice(memoryStorage())).toBe(LONG_FORM_DEFAULT);
    expect(loadLongFormChoice(null)).toBe(LONG_FORM_DEFAULT);
@@ -75,7 +75,7 @@ describe.skipIf(!available)("Silero VAD on the Transformers.js runtime (real mod
  });
 });

-describe("createWhisperDeps decode with the names-and-terms hint (D-WS2a × D-WS3a)", () => {
+describe("createWhisperDeps decode with the names-and-terms hint (D-048 × D-049)", () => {
  /** A fake ASR pipeline that records the generate options of each call. */
  function fakeAsr() {
    const calls: Record<string, unknown>[] = [];
@@ -106,7 +106,7 @@ describe("shouldWarnBeforeLoad", () => {
  });
 });

-describe("desktop-only models (D-WS7a)", () => {
+describe("desktop-only models (D-051)", () => {
  const parakeet = MODELS.find((model) => model.id === PARAKEET_MODEL_ID)!;
  const phone = { userAgent: ANDROID_UA, maxTouchPoints: 5, deviceMemoryGB: 8 };
  const ipad = { userAgent: IPAD_DESKTOP_UA, maxTouchPoints: 5, deviceMemoryGB: undefined };
@@ -2,7 +2,7 @@ import { describe, expect, test } from "bun:test";
 import { fetchModelFile, type ModelFileCacheStorage } from "../src/worker/modelFiles";
 import type { LoadProgressMessage } from "../src/worker/protocol";

-// Cache-API-backed weight download for the Parakeet path (D-WS7a, D-005).
+// Cache-API-backed weight download for the Parakeet path (D-051, D-005).

 const URL_A = "https://huggingface.co/org/model/resolve/rev/encoder.onnx";

@@ -39,7 +39,7 @@ describe("MODELS catalog", () => {
      "onnx-community/whisper-medium-ONNX",
      "onnx-community/whisper-large-v3-turbo",
      `${DANISH_MODELS_HF_OWNER}/whisper-small-danish-ONNX`,
-      // Not Whisper: the D-WS7a Parakeet engine path, last so it never
+      // Not Whisper: the D-051 Parakeet engine path, last so it never
      // becomes anyone's accidental default.
      "ysdede/parakeet-tdt-0.6b-v3-onnx",
    ]);
@@ -80,7 +80,7 @@ describe("MODELS catalog", () => {
  });

  test("every Whisper model maps to a faster-whisper name for the server engine (D-013)", () => {
-    // The Danish conversions (D-WS6a) have no faster-whisper alias; their
+    // The Danish conversions (D-046) have no faster-whisper alias; their
    // server-engine name is a Hub repo id — covered by the Danish suite below.
    for (const model of WHISPER_MODELS.filter((m) => !m.id.startsWith(`${DANISH_MODELS_HF_OWNER}/`))) {
      // The onnx-community ids and faster-whisper's aliases share the same
@@ -98,7 +98,7 @@ describe("MODELS catalog", () => {
  });
 });

-describe("Danish fine-tune conversions (D-WS6a)", () => {
+describe("Danish fine-tune conversions (D-046)", () => {
  const owner = `${DANISH_MODELS_HF_OWNER}/`;
  const danish = MODELS.filter((model) => model.id.startsWith(owner));

@@ -131,7 +131,7 @@ describe("Danish fine-tune conversions (D-WS6a)", () => {
    expect(smallDa!.decoderParamsM).toBe(small.decoderParamsM);
  });

-  test("default to the dtypes measured on the converted files (D-WS6a)", () => {
+  test("default to the dtypes measured on the converted files (D-046)", () => {
    // da_festuge on onnxruntime-node: q4 matched or beat q8 and was faster for
    // all three, small included; fp16 encoders did not overflow (D-022 check).
    for (const model of danish) {
@@ -141,7 +141,7 @@ describe("Danish fine-tune conversions (D-WS6a)", () => {
        decoder_model_merged: "q4",
      });
      expect(resolveDefaultDtype(model.id, "webnn")).toBe("fp16");
-      expect(model.defaults.wasm.note).toContain("D-WS6a");
+      expect(model.defaults.wasm.note).toContain("D-046");
    }
  });
 });
@@ -270,7 +270,7 @@ describe("WHISPER_LANGUAGES", () => {
  });
 });

-describe("canTranslate (D-WS4a)", () => {
+describe("canTranslate (D-044)", () => {
  test("Whisper translate is offered only where OpenAI trained it", () => {
    const translating = MODELS.filter((model) => model.canTranslate).map((model) => model.id);
    // large-v3-turbo was not trained for translation (OpenAI model card);
@@ -290,7 +290,7 @@ describe("canTranslate (D-WS4a)", () => {
  });
 });

-describe("taskForModel (D-WS4a)", () => {
+describe("taskForModel (D-044)", () => {
  test("a stale translate resets to transcribe on a model that cannot translate", () => {
    expect(taskForModel("onnx-community/whisper-large-v3-turbo", "translate")).toBe("transcribe");
    expect(taskForModel("onnx-community/whisper-small.en", "translate")).toBe("transcribe");
@@ -306,7 +306,7 @@ describe("taskForModel (D-WS4a)", () => {
  });
 });

-describe("Parakeet catalog entry (D-WS7a)", () => {
+describe("Parakeet catalog entry (D-051)", () => {
  const parakeet = getModel(PARAKEET_MODEL_ID)!;

  test("family defaults to whisper when absent", () => {
@@ -69,7 +69,7 @@ The worker now caps WASM sessions at `graphOptimizationLevel: "basic"`
 outside the browser. **Re-run this step with Backend = WASM to confirm in a
 real browser.**

-**Update (2026-10-30, D-WS1c):** the Transformers.js 4.3.2 bump brings an
+**Update (2026-10-30, D-043):** the Transformers.js 4.3.2 bump brings an
 onnxruntime-web build without that bug, so the cap is gone and WASM sessions
 run at the default level again. Re-run this step with Backend = WASM (Whisper
 Tiny and Tiny.en) to confirm the session still loads in a real browser.
@@ -111,7 +111,7 @@ Result: Works.

 Result: Works

-## 5. Transcript translation offline (D-WS4b)
+## 5. Transcript translation offline (D-045)

 1. Online: transcribe anything, open **Translate transcript…**, pick a
   direction, click **Translate**. The hint said "Downloads a translation
@@ -129,7 +129,7 @@ Result: Works

 Result: _pending (needs a real browser)_

-## 6. "Skip silence" (long-form pipeline, D-WS3a/D-WS3b)
+## 6. "Skip silence" (long-form pipeline, D-049/D-050)

 1. Online, open **Advanced**, tick **Skip silence in uploaded files**, and
   transcribe a file. DevTools → Network shows one request for
@@ -15,7 +15,7 @@ import {
 import { wordErrorRate } from "./dsp";

 /**
- * Real-model accuracy of the Parakeet engine path (D-WS7a), opt-in:
+ * Real-model accuracy of the Parakeet engine path (D-051), opt-in:
 *
 *   PARAKEET_REAL=1 bun test tests/parakeet-real-model.test.ts
 *
@@ -12,7 +12,7 @@ import {
 } from "../src/lib/parakeet";
 import { transcriptToSrt, transcriptToTimestampedTxt } from "../src/lib/export";

-// Pure Parakeet engine logic (D-WS7a) — no model, no runtime.
+// Pure Parakeet engine logic (D-051) — no model, no runtime.

 const RATE = 16_000;

@@ -35,7 +35,7 @@ function tokens(output: unknown): number[] {
  return (output as Tensor).tolist()[0]!.map(Number);
 }

-describe("installPromptSupport (D-WS2a)", () => {
+describe("installPromptSupport (D-048)", () => {
  test("without a prompt kwarg, generate is called exactly as before", async () => {
    const model = fakeModel([50258, 50259, 50360]);
    installPromptSupport(model);
@@ -117,7 +117,7 @@ describe("buildPromptPrefix", () => {
    ]);
  });

-  test("a huge hint stays within the 120-token budget (D-WS2a)", () => {
+  test("a huge hint stays within the 120-token budget (D-048)", () => {
    const prefix = buildPromptPrefix(tokenizer, "Maxwell, ".repeat(500));
    expect(prefix.length).toBeLessThanOrEqual(121);
  });
@@ -83,7 +83,7 @@ describe("isUiToWorkerMessage", () => {
  });
 });

-describe("choosePipeline (D-WS3a/D-WS3b)", () => {
+describe("choosePipeline (D-049/D-050)", () => {
  test("long-form only when asked for and the VAD model loaded", () => {
    expect(choosePipeline({ longForm: true, vadReady: true })).toBe("longform");
    // Offline / never downloaded: the old path, never an error (Review Focus 3).
@@ -248,7 +248,7 @@ describe("deviceAttempts (D-004, D-017)", () => {
  });
 });

-describe("parakeetDeviceAttempts (D-WS7a)", () => {
+describe("parakeetDeviceAttempts (D-051)", () => {
  const everything = { webnn: true, webgpu: true, webgpuF16: true };

  test("auto never picks an accelerator for Parakeet", () => {
@@ -322,7 +322,7 @@ describe("TranscribeProgressTracker", () => {
    expect(tracker.finalizeWindow()).toBeNull(); // 20 clamps to 8, not an advance
  });

-  describe("forWindows (variable-length VAD chunks, D-WS3a)", () => {
+  describe("forWindows (variable-length VAD chunks, D-049)", () => {
    // Two speech chunks with a skipped 15 s silence between them: [0–20 s]
    // and [35–60 s] of a 70 s file.
    const windows = [
@@ -22,7 +22,7 @@ FW_INTEGRATION=1 bun test tests/transcribe.test.ts

 Runs the real sidecar end to end (spawn → ready handshake → `tiny.en` load →
 transcribe → done), then the first 60 s of `tests/audio_files/en_alien.opus`
-(needs `ffmpeg`) in both decoding modes (D-WS5a): `file` (faster-whisper's
+(needs `ffmpeg`) in both decoding modes (D-047): `file` (faster-whisper's
 `BatchedInferencePipeline` + Silero VAD, what `POST /api/transcribe` uses) and
 `live` (sequential, what server-live passes use). First run downloads ~75 MB
 of weights from the Hub. Point `SIDECAR_PYTHON` at a virtualenv's python if
@@ -52,11 +52,11 @@ burst, segments in bursts — and cancels a batched run after its first burst.
 5. Transcribe the same clip through the server engine.

 - [ ] Segments stream in as JSON lines (not one blob at the end). With the
-      batched pipeline (D-WS5a) they arrive in bursts — one burst per 8
+      batched pipeline (D-047) they arrive in bursts — one burst per 8
      speech chunks (≈ up to 4 min of audio) — with blank heartbeat lines
      every 5 s in between; on a clip under ~30 s of speech, everything
      arrives as a single segment at the end. Each segment covers one VAD
-      speech chunk (up to 30 s), not one sentence — expected (D-WS5a).
+      speech chunk (up to 30 s), not one sentence — expected (D-047).
 - [ ] Cancelling a long server transcription (abort / page refresh) stops
      the sidecar within one burst; the next request then starts normally.
 - [ ] Both transcripts agree in content (minor punctuation/casing drift
@@ -125,7 +125,7 @@ describe("createLiveWebSocket", () => {
    // wait for the payload itself, or a loaded machine races past it.
    await waitFor(() => sidecar.audioPayloads.length > 0);
    expect(sidecar.requests[0]).toEqual(
-      // Live passes are short: sequential decoding, no batching/VAD (D-WS5a).
+      // Live passes are short: sequential decoding, no batching/VAD (D-047).
      expect.objectContaining({ model: "base", language: "da", task: "translate", mode: "live" }),
    );
    // The bytes written to the temp file are exactly the committed segment's
@@ -142,7 +142,7 @@ describe("createLiveWebSocket", () => {
    await waitFor(async () => (await readdir(tempDir)).length === 0);
  });

-  test("every pass of a session carries the session's vocabulary hint (D-WS2a)", async () => {
+  test("every pass of a session carries the session's vocabulary hint (D-048)", async () => {
    const tempDir = await makeTempDir();
    const sidecar = fakeSidecar();
    const handler = createLiveWebSocket({ sidecar });
@@ -168,7 +168,7 @@ describe("parseLanguageAndTask", () => {
  });
 });

-describe("parseLanguageAndTask: vocabulary hint (D-WS2a)", () => {
+describe("parseLanguageAndTask: vocabulary hint (D-048)", () => {
  const parse = (query: string) => parseLanguageAndTask(new URL(`http://localhost/x${query}`));

  test("normalizes the hint like the worker does", () => {
@@ -357,7 +357,7 @@ describe("POST /api/transcribe", () => {
  });

  test("heartbeat blank lines keep a quiet stream alive before the first segment (D-021)", async () => {
-    // The batched pipeline (D-WS5a) runs VAD over the whole file and decodes a
+    // The batched pipeline (D-047) runs VAD over the whole file and decodes a
    // full batch before its first segment, so the quiet stretch is longer.
    const handlers = createTranscribeHandlers({
      sidecar: fakeSidecar({ delayMs: 200 }),
@@ -375,7 +375,7 @@ describe("POST /api/transcribe", () => {
    expect(text).toContain('"type":"result"');
  });

-  test("passes the vocabulary hint to the sidecar; 400 for an over-long one (D-WS2a)", async () => {
+  test("passes the vocabulary hint to the sidecar; 400 for an over-long one (D-048)", async () => {
    const sidecar = fakeSidecar();
    const handlers = createTranscribeHandlers({
      sidecar,
@@ -477,7 +477,7 @@ describe("createSidecar (fake python sidecar)", () => {
    expect(segments[1]!.text).toContain("model=ok bytes=12 lang=da task=translate");
    expect(segments[1]!.text).not.toContain("prompt=");

-    // The vocabulary hint crosses the JSON-lines protocol intact (D-WS2a).
+    // The vocabulary hint crosses the JSON-lines protocol intact (D-048).
    const prompted: SidecarSegment[] = [];
    await sidecar.transcribe(
      { model: "ok", audioPath, prompt: "Århus, Voices of the Void" },
@@ -554,11 +554,11 @@ describe("createSidecar (degradation, D-013)", () => {

 // ---------------------------------------------------------------------------
 // The real server-py/sidecar.py against a stubbed faster_whisper/numpy
-// (tests/fixtures/stub_fw): request routing, model cache, cancel (D-WS5a).
+// (tests/fixtures/stub_fw): request routing, model cache, cancel (D-047).
 // `-S` keeps real site-packages out, so this runs the same everywhere.
 // ---------------------------------------------------------------------------

-describe("server-py/sidecar.py (stubbed faster-whisper, D-WS5a)", () => {
+describe("server-py/sidecar.py (stubbed faster-whisper, D-047)", () => {
  const STUB_DIR = join(import.meta.dir, "fixtures/stub_fw");
  const stubbedSidecar = () =>
    createSidecar(["env", `PYTHONPATH=${STUB_DIR}`, "python3", "-S", REAL_SIDECAR]);
@@ -590,7 +590,7 @@ describe("server-py/sidecar.py (stubbed faster-whisper, D-WS5a)", () => {
    expect(explicit.texts[0]).toContain("batched model=m1");
    expect(explicit.texts[0]).toContain("vad_filter=True");
    expect(explicit.texts[0]).toContain("batch_size=8");
-    // Tuned Silero VAD: the default (0.5 / 400 ms) dropped quiet speech (D-WS5a).
+    // Tuned Silero VAD: the default (0.5 / 400 ms) dropped quiet speech (D-047).
    expect(explicit.texts[0]).toContain("vad_parameters={'threshold': 0.2, 'speech_pad_ms': 800}");
    expect(explicit.texts[0]).toContain("language=da");
    expect(explicit.texts[0]).toContain("task=transcribe");
@@ -599,7 +599,7 @@ describe("server-py/sidecar.py (stubbed faster-whisper, D-WS5a)", () => {
    expect(implicit.texts[0]).toContain("batched model=m1");
  });

-  test("the vocabulary hint reaches the batched pipeline as text, and live passes as ids (D-WS2a, D-WS5a)", async () => {
+  test("the vocabulary hint reaches the batched pipeline as text, and live passes as ids (D-048, D-047)", async () => {
    const sidecar = stubbedSidecar();
    const file = await run(sidecar, { model: "m1", mode: "file", prompt: "Maxwell, Voices of the Void" });
    expect(file.texts).toHaveLength(1);
@@ -611,7 +611,7 @@ describe("server-py/sidecar.py (stubbed faster-whisper, D-WS5a)", () => {
    expect(live.texts[0]).toContain("initial_prompt=[32, 77, 97, 120, 119, 101, 108, 108, 46]");
  });

-  test("live mode stays sequential without VAD (short passes, D-WS5a)", async () => {
+  test("live mode stays sequential without VAD (short passes, D-047)", async () => {
    const sidecar = stubbedSidecar();
    const { texts } = await run(sidecar, { model: "m1", mode: "live", task: "translate" });
    expect(texts[0]).toContain("seq model=m1");
@@ -681,7 +681,7 @@ describe.skipIf(!process.env.FW_INTEGRATION)("real faster-whisper sidecar", () =
  );

  test(
-    "file mode (batched + VAD) and live mode both transcribe real speech (D-WS5a)",
+    "file mode (batched + VAD) and live mode both transcribe real speech (D-047)",
    async () => {
      const sidecar = createSidecar();
      // First 60 s of en_alien as raw float32 16 kHz mono — what the browser uploads.
@@ -718,7 +718,7 @@ describe.skipIf(!process.env.FW_INTEGRATION)("real faster-whisper sidecar", () =
  }

  test(
-    "a long file streams in bursts behind heartbeats over HTTP (D-021, D-WS5a)",
+    "a long file streams in bursts behind heartbeats over HTTP (D-021, D-047)",
    async () => {
      const handlers = createTranscribeHandlers({
        sidecar: createSidecar(),
@@ -756,7 +756,7 @@ describe.skipIf(!process.env.FW_INTEGRATION)("real faster-whisper sidecar", () =
      expect(segmentLines.at(-1)!.at - segmentLines[0]!.at).toBeGreaterThan(100);
      expect(lines.at(-1)!.line).toContain('"type":"result"');
      console.log(
-        "[D-WS5a] segment arrival (ms):",
+        "[D-047] segment arrival (ms):",
        segmentLines.map(({ at }) => Math.round(at)).join(" "),
      );
    },
@@ -2,7 +2,7 @@ import { describe, expect, test } from "bun:test";
 import { renderToStaticMarkup } from "react-dom/server";
 import { TranscriptionOptions } from "../src/components/TranscriptionOptions";

-// D-WS4a: Whisper's own "Translate to English" only where the model can do it.
+// D-044: Whisper's own "Translate to English" only where the model can do it.
 function render(canTranslate: boolean): string {
  return renderToStaticMarkup(
    <TranscriptionOptions
@@ -3,7 +3,7 @@ import { existsSync, readFileSync } from "node:fs";
 import { join } from "node:path";

 /**
- * Tripwires for the patched, pinned Transformers.js runtime (D-WS1a).
+ * Tripwires for the patched, pinned Transformers.js runtime (D-040).
 * `@huggingface/transformers` carries a `bun patch` (D-027/D-028) keyed by
 * its exact version; a bump that forgets to regenerate the patch would leave
 * the app working but silently bring back the mobile OOM. These checks fail
@@ -48,22 +48,22 @@ describe("@huggingface/transformers runtime", () => {
    expect(body).toContain("for (const name of Object.keys(names))");
  });

-  test("Whisper generate still takes the seek loop when max_new_tokens is unset (D-021, D-WS1b)", () => {
+  test("Whisper generate still takes the seek loop when max_new_tokens is unset (D-021, D-042)", () => {
    // The worker's `max_new_tokens: 440` exists only to dodge this branch.
    // If upstream changes the condition, re-measure D-021 before touching it.
    const source = read(join(PKG_DIR, "src/models/whisper/modeling_whisper.js"));
    expect(source).toContain("if (generation_config.return_timestamps && !kwargs.max_new_tokens)");
  });

-  test("onnxruntime-web is the build D-WS1c re-tested 8-bit decoders on at the default optimization level", () => {
+  test("onnxruntime-web is the build D-043 re-tested 8-bit decoders on at the default optimization level", () => {
    // D-014's "basic" cap was dropped because this build loads the q8 merged
    // decoders at every graphOptimizationLevel. A different build needs the
-    // D-WS1c session-creation check re-run before it ships.
+    // D-043 session-creation check re-run before it ships.
    const ortWeb = JSON.parse(read(join(ROOT, "node_modules/onnxruntime-web/package.json"))).version;
    expect(ortWeb).toBe("1.31.0-dev.20260914-8d85527a0");
  });

-  test("worker leaves WASM sessions at the default graph optimization level (D-WS1c)", () => {
+  test("worker leaves WASM sessions at the default graph optimization level (D-043)", () => {
    const worker = read(join(ROOT, "src/worker/inference.worker.ts"));
    expect(worker).not.toContain("graphOptimizationLevel");
  });
@@ -9,7 +9,7 @@ import {
 import { loadOpusMt, type OpusMtTranslator } from "../src/worker/opusMt";

 /**
- * Real Opus-MT translation on onnxruntime-node (D-WS4b), opt-in:
+ * Real Opus-MT translation on onnxruntime-node (D-045), opt-in:
 *
 *   MT_INTEGRATION=1 flock /tmp/whisper-heavy.lock bun test tests/translate-real-model.test.ts
 *
@@ -2,7 +2,7 @@ import { describe, expect, test } from "bun:test";
 import { renderToStaticMarkup } from "react-dom/server";
 import { TranslatePanel, translateStatusText } from "../src/components/TranslatePanel";

-// D-WS4b: the panel's copy, before any worker exists (static render).
+// D-045: the panel's copy, before any worker exists (static render).
 const danish = { text: "Det er ikke noget med at vi skal på ferie.", chunks: [] };

 function render(downloaded: string[] = [], hint?: string) {
@@ -12,7 +12,7 @@ import {
 } from "../src/lib/translation";
 import type { TranscriptChunk } from "../src/worker/protocol";

-// Pure planning for transcript translation (D-WS4b) — no model, no network.
+// Pure planning for transcript translation (D-045) — no model, no network.

 describe("translateUnits", () => {
  test("translates every piece in order, batched, and rejoins per unit", async () => {
@@ -2,7 +2,7 @@ import { describe, expect, test } from "bun:test";
 import { join } from "node:path";
 import { ortAliasPlugin } from "../scripts/ortAlias";

-// D-WS7a: the Parakeet path must not bring a second onnxruntime into the
+// D-051: the Parakeet path must not bring a second onnxruntime into the
 // inference worker. parakeet.js depends on its own onnxruntime-web (1.24.1);
 // the app runs the one Transformers.js pins. Unminified bundles carry a
 // `// node_modules/<path>` comment per bundled module, so count runtimes.