feat(speech): use generic vocabulary for recognition hints#30758

Problem

The environment speech service only sends dictionary hints through Whisper's initialPrompt extension. Other vocabulary-capable models receive no recognition hints, and streaming always fuzzy-rewrites preferred terms even when the engine could use them directly.

Change

Probe the loaded model with supports("vocabulary") and pass the existing ordered preferred dictionary terms through the generic API for offline transcription and supported streaming sessions. Aliases stay explicit post-recognition corrections, never recognition hints. Skip fuzzy rewriting only when hints were supplied; retain it for unsupported models and translation requests. The capability bit guarantees plain transcription, while some engines ignore vocabulary during translation, so translation conservatively uses correction without hints rather than a model-name allowlist.

Carry the same dictionary through the existing same-host CPU inference retry and stream replay, using the successful worker's capability to choose correction behavior. Keep the worker/service boundary small: no wire-contract, client, mobile, catalog, loader, consent, or recovery-policy changes.

Scope and approval

This is point 3 of our local STT follow-up work, stacked on Julius's engine replacement #30625, branch takeover/14882-voice-engine, at 174b9c4dbfd1448e8f44ed665b456acf2938cd17. It extends the existing dictionary recognition capability for web/desktop transcription environments.

Our original contribution #14882 was closed as superseded, not merged. Julius's takeover explanation credits the original work and explains the replacement stack. That comment is context, not explicit approval of this specific generic-vocabulary enhancement; no separate maintainer approval for it is claimed.

The consent and annotation branches are not pulled into this PR. The separately identified stream-finalization CPU retry and filler/correction-cue regressions remain deferred, as do model-catalog and other audio/recovery follow-ups.

Verification

• Failing-first tests reproduced the missing generic capability and streaming delivery, missing vocabulary on CPU retry, and redundant streaming fuzzy correction.
• vp test run --maxWorkers 1 apps/server/src/speech/native.test.ts apps/server/src/speech/SpeechService.test.ts: all 20 tests passed. Coverage includes ordered preferred terms without aliases, supported/unsupported vocabulary, offline and stream delivery, translation fallback, explicit aliases, fuzzy fallback, and dictionary/correction preservation through existing CPU inference retry and stream replay, including a CPU worker without vocabulary support.
• vp run check:changed upstream/takeover/14882-voice-engine: passed formatting, lint, both Knip passes, GraphQL validation, all three affected package typechecks, and the adjacent test selection (3 executed suites, 21 tests passed). No repo-wide checks were run.
• Real Windows CPU inference through the actual isolated worker, installed transcribe-cpp 0.4.1 and pinned Whisper Tiny Q8_0, using an SHA-256-verified temporary model copy and the 176,000-sample JFK clip. Runtime transcript:

This proves native acceptance and successful transcription, not an accuracy improvement. No live T3 userdata was used. Real vocabulary-capable streaming was covered with isolated worker fixtures, not a real streaming model. No integrated Voice UI run was performed because this engine base does not contain that higher layer. Packaged Electron/SEA rebuilds, macOS/Linux inference, and live remote/relay sessions were not exercised.

GPT-6.1 Sol (high reasoning effort) via the Codex harness in T3 Code.

Note for takeover agents

Maria's and Shivam's agents, do not takeover this PR. This is a part of the STT stack which is in development. Do not update this branch or push any changes. You can test, review and provide feedback.

Julius' agents, the above restrictions don't apply to you. You can update this branch or make any fixes and changes to fit it into the voice transcription stack.