feat(stt): transcribe with omp's own speech-to-text models, including a ChatGPT/Codex subscription

[andrebrait]

Problem

Voice input needed an OpenAI-compatible STT endpoint (OMP_WEB_STT_ENDPOINT): an API key or a separately hosted STT server. Users with a ChatGPT/Codex subscription already logged in to omp could not dictate with it.

Changes

• Transcribe with an omp model. Leave OMP_WEB_STT_ENDPOINT unset and set OMP_WEB_STT_MODEL to an omp speech-to-text model, such as openai-codex/codex-transcribe (ChatGPT subscription, added in can1357/oh-my-pi#15391). An endpoint, when set, still wins, so existing setups are unchanged.
• omp keeps the credentials. Jobs go through omp auth-gateway stdio (lib/omp/auth-gateway.ts): one shared child, started on first use and stopped after 5 minutes idle, sending base64 JSON requests. omp-web never reads omp's credential store. A crashed child rejects its in-flight jobs and the next job respawns it; a child that never reports ready is killed after 60 seconds.
• Size cap. The gateway limits a JSON body to 25 MB and base64 grows audio by 4/3, so this path rejects recordings over 18 MB with a clear error instead of an opaque 413.
• The 501 message names both variables. README (en/ja/zh-CN), CHANGELOG and AGENTS.md describe the setting.

Verification

• npm test: 1356 pass, 0 fail. lib/stt-omp-gateway.test.mjs runs real jobs against a stub omp that speaks the stdio protocol: a transcript, an upstream error, a crash (error carries omp's stderr, next job respawns the gateway), and the format following the browser's file name (webm, mp4). The 501 test now clears both variables.
• tsc --noEmit clean; eslint has no errors.
• Live, against a real Codex subscription with the omp branch from feat(ai): transcribe speech with a ChatGPT/Codex subscription can1357/oh-my-pi#15391: the job routes with OMP_WEB_STT_MODEL=openai-codex/codex-transcribe transcribed the Whisper JFK clip correctly, 5.9 s cold (gateway start) and 1.1 s warm.
• Endpoint probes: WebM (Chrome), MP4 (Safari) and Ogg/Opus (Firefox) recordings all transcribe.

Needs an omp release that includes can1357/oh-my-pi#15391 for the Codex model; other omp STT models (for example openai/whisper-1 with an OpenAI key in omp) also need its auth-gateway stdio selector fix.