PR: https://git.cafe/pingdotgg/t3code/pulls/30758
Head: 625580da65c0c0b43e47ebb4f5d6329d96fd2f17
Linux x64 GNU; transcribe-cpp 0.4.1; CPU inference.
Only one task dev server ran at a time. Isolated home: /home/user/t3code/.t3.

Main baseline: github/main 8b44f42b14, separate /home/user/main worktree, vp i, background vp run dev.
Authenticated web-origin GET /api/voice/status -> 404 Not Found.
No Voice UI or transcription endpoint exists on this main revision.

PR: vp i; background vp run dev --home-dir /home/user/t3code/.t3.
Authenticated GET /api/voice/status -> 200, supported=true.
Models downloaded and selected through /api/voice/models/download and /select.
Actual audio: https://raw.githubusercontent.com/ggml-org/whisper.cpp/master/samples/jfk.wav
Converted with ffmpeg to 16 kHz mono Float32 PCM, no synthetic audio.
POST /api/voice/transcribe uses four-byte little-endian metadata length, JSON options, then PCM.

Native runtime capability probes (real model load, followed by disposal):
Whisper Tiny: backend=CPU, supportsVocabulary=true, supportsStreaming=false, supportsTranslation=true.
Qwen3-ASR 0.6B: backend=CPU, supportsVocabulary=true, supportsStreaming=false, supportsTranslation=false.
Parakeet Unified EN 0.6B: backend=CPU, supportsVocabulary=false, supportsStreaming=true, supportsTranslation=false.
Canary 180M Flash: backend=CPU, supportsVocabulary=false, supportsStreaming=false, supportsTranslation=true.

Qwen request dictionary: [{term:"Amaricans",aliases:[]},{term:"T3 Code",aliases:["fellow"]}]
Language=en, translation=false, filler removal=false. Same audio and metadata for both revisions.
Engine base 174b9c4dbfd1448e8f44ed665b456acf2938cd17, running server:
200 {"text":"And so, my T3 Code Amaricans, ask not what your country can do for you; ask what you can do for your country."}

PR head, restarted and verified again:
200 {"text":"And so, my T3 Code Americans, ask not what your country can do for you; ask what you can do for your country."}

Qwen without dictionary:
200 {"text":"And so, my fellow Americans, ask not what your country can do for you; ask what you can do for your country."}

Whisper Tiny on PR, HTTP 200 each:
Plain: And so my fellow Americans ask not what your country can do for you ask what you can do for your country.
Preferred vocabulary [Americans, country, fellow citizens]: and so my fellow Americans ask not what your country can do for you ask what you can do for your country.
Explicit alias Americans -> T3 Code: And so my fellow T3 Code ask not what your country can do for you ask what you can do for your country.
Translation request (language=fr, translateToEnglish=true) with same alias: And so my fellow T3 Code ask not what your country can do for you, ask what you can do for your country.
The translation check proves request-path compatibility and alias correction, not French translation accuracy.

Parakeet real /ws/voice stream: start -> ready -> 8 ordered updates -> finished.
Dictionary [{term:"T3 Code",aliases:["Americans"]},{term:"Country",aliases:[]}].
Final: And so, my fellow T3 Code, ask not what your Country can do for you. Ask what you can do for your Country
Aliases and dictionary spelling visible in partial updates too. Full events attached separately.

Playwright from apps/web dependencies, /usr/bin/google-chrome:
Created project /home/user/speech-verification (real git repo) using Add project > Local folder.
Created thread and selected Codex GPT-6.1-Sol in model picker.
Submitted two prompts through composer. Both real provider turns completed:
"speech verification provider path works."
"neighboring chat behavior still works."
Provider screenshot and trimmed recording capture the actual app.

Blacksmith:
vp test run apps/server/src/speech/native.test.ts apps/server/src/speech/SpeechService.test.ts apps/server/src/speech/http.test.ts apps/server/src/speech/stream.test.ts
4 suites, 29 tests passed.
vp run check:changed HEAD^ (engine base fetched into remote runner first)
format, lint, both knip passes, GraphQL, 3-package typecheck, 3 suites / 21 tests passed.
No source changes, commits or pushes were needed.

Not verified: Voice composer/dictionary UI (absent from this stacked head), native mobile/desktop packaging, macOS/Windows, remote/relay/tunnel, real GPU-to-CPU recovery (no GPU), vocabulary-capable native streaming (the real streaming model tested reports supportsVocabulary=false), recognition accuracy improvement. Recovery/capability matrices passed in existing automated tests only.