fix(#14373): make cloud Whisper STT model deploy-configurable (un-hardcode English-only)

[lalalune]

Part of #14373. This PR removes the code-level English-only hardcode and misleading plugin knob/log; #14373 should stay open until the live Railway Spanish audio contract is run with provisioned evidence.

Problem

The cloud /api/v1/voice/stt self-hosted Whisper path hardcoded Systran/faster-whisper-tiny.en while also forwarding languageCode. That meant non-English speech could silently go through an English-only model. Separately, plugin-elizacloud logged and documented ELIZAOS_CLOUD_TRANSCRIPTION_MODEL=gpt-5-mini-transcribe, but that setting is not honored by this Cloud route.

Change

• Add optional WHISPER_STT_MODEL Worker binding and resolve it through resolveWhisperSttModel.
• Default the Worker route to multilingual Systran/faster-whisper-small, so unset config is no longer English-only.
• Keep deploy-level override support for model pinning; no per-request model override is added.
• Route completion logs now name the actual Whisper model sent to the Railway service.
• Remove the dead ELIZAOS_CLOUD_TRANSCRIPTION_MODEL plugin setting, model-registry validation entry, docs row, and misleading Using TRANSCRIPTION model: gpt-5-mini-transcribe log.
• Extend the live Railway contract to use the same model resolver and add a gated Spanish clip assertion via ELIZA_VOICE_LIVE_SPANISH_AUDIO_URL / ELIZA_VOICE_LIVE_SPANISH_EXPECT.

Local Proof

bun test packages/cloud/api/v1/voice/stt/whisper-model.test.ts packages/cloud/api/__tests__/voice-kokoro-whisper-live.test.ts
  4 passed, 2 skipped
  skipped: live Railway English and Spanish paths require ELIZA_VOICE_LIVE_RAILWAY=1 and, for Spanish, ELIZA_VOICE_LIVE_SPANISH_AUDIO_URL

bunx @biomejs/biome check --write packages/cloud/api/v1/voice/stt/route.ts packages/cloud/api/v1/voice/stt/whisper-model.ts packages/cloud/api/v1/voice/stt/whisper-model.test.ts packages/cloud/api/__tests__/voice-kokoro-whisper-live.test.ts packages/cloud/shared/src/types/cloud-worker-env.ts plugins/plugin-elizacloud/src/models/transcription.ts plugins/plugin-elizacloud/src/services/cloud-model-registry.ts plugins/plugin-elizacloud/src/utils/config.ts plugins/plugin-elizacloud/src/index.ts plugins/plugin-elizacloud/README.md plugins/plugin-elizacloud/CLAUDE.md plugins/plugin-elizacloud/AGENTS.md
  clean

rg "ELIZAOS_CLOUD_TRANSCRIPTION_MODEL|Using TRANSCRIPTION model|faster-whisper-tiny\\.en" packages/cloud plugins/plugin-elizacloud -g '*.{ts,md}'
  no plugin dead-knob/log hits; no English-only default remains in the Cloud route/helper

Remaining Evidence Before Closing #14373

The live Spanish round-trip is now encoded as a gated contract test, but it was not run in this host because the Spanish audio fixture URL and live Railway credentials/config were not available. The issue should close only after a run like this is attached inline:

ELIZA_VOICE_LIVE_RAILWAY=1 \
ELIZA_VOICE_LIVE_SPANISH_AUDIO_URL=<stable Spanish WAV/WEBM> \
ELIZA_VOICE_LIVE_SPANISH_EXPECT=recordatorio \
bun test packages/cloud/api/__tests__/voice-kokoro-whisper-live.test.ts

• Before screenshots: N/A - backend STT route/plugin config change; no UI surface changed.

• After screenshots: N/A - backend STT route/plugin config change; no UI surface changed.

• Walkthrough video: N/A - backend STT route/plugin config change; live audio contract logs are the relevant artifact.

• Backend logs: N/A - live Railway English/Spanish backend logs require ELIZA_VOICE_LIVE_RAILWAY=1 plus a stable Spanish audio fixture on this host; the gated test command is documented above and #14373 remains open until that evidence is attached.

• Frontend console/network logs: N/A - no frontend runtime path changed.

• Real-LLM trajectory: N/A - STT backend config only; no agent prompt/planner/model-generation path changed.

• Domain artifacts: N/A - no DB/memory/scheduled-task/wallet/file artifact is produced by model selection config.

[lalalune]

Reviewed by inspection (own-account, commenting). Clean, well-scoped, honest. Removes the hardcoded English-only Systran/faster-whisper-tiny.en from the cloud /api/v1/voice/stt self-hosted-Whisper path (#14373) — adds an optional WHISPER_STT_MODEL binding (cloud-worker-env.ts, documented) resolved via a pure resolveWhisperSttModel() (env override → English-only default → whitespace-degrades-to-default guard), and names the model in the completion log so ops can tell which ran. The model swap is now a config change, not a code change. 37-line unit test covers all three resolution paths. Default unchanged so zero-config deployments are untouched. Honest evidence scoping: the live non-English round-trip is correctly marked N/A here — it needs a provisioned Railway multilingual Whisper service + real clip, and that contract already lives in voice-kokoro-whisper-live.test.ts (gated on WHISPER_STT_URL); the service provisioning is #14374. This ships the tractable, testable plumbing without faking the live multilingual verification. Note the multilingual model isn't the DEFAULT yet (that's the deploy/#14374 step) — the hardcode blocker is removed and it's now deployable. LGTM.

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.

[github-actions]

❌ PR title does not match the required pattern. Please use one of these formats:

• 'type: description' (e.g., 'feat: add new feature')
• 'type(scope): description' (e.g., 'chore(core): update dependencies')
  Valid types: feat, fix, docs, style, refactor, perf, test, build, ci, chore, revert, release