Revert native audio transcription support

[Schwartz10]

Summary

• Revert the squash-merged native audio/STT PR ([codex] Add native audio input support #2, commit 49566b3)
• Remove the /v1/audio/transcriptions utility endpoint and Codex subscription STT forwarding path
• Remove the associated capability advertisement, runtime-provider fallback knob, and tests

Why

The subscription-based ChatGPT/Codex transcription path does not work for this repository, and keeping the endpoint creates a misleading integration surface.

Validation

• .venv/bin/python -m pytest tests/gateway/test_api_server.py tests/gateway/test_api_server_multimodal.py tests/gateway/test_session_api.py tests/hermes_cli/test_runtime_provider_resolution.py tests/agent/test_models_dev.py tests/run_agent/test_run_agent_multimodal_prologue.py
• Result: 425 passed

[Schwartz10]

@claude can you review this PR? We're reverting transcription from subscription support. i dont want any dead code hanging around, or legacy feature suppport for the existing PR

[claude]

Claude finished @Schwartz10's task in 2m 36s —— View job

---

Review: Revert native audio transcription support

• Gather context (read diff vs origin/williamos/main, compared against original squash-merge commit 49566b3c4)
• Review changed files for leftover dead code / legacy references
• Check tests still reference removed feature
• Post review feedback

Summary

I diffed this branch against origin/williamos/main and also compared it line-by-line against the original squash-merged commit (49566b3c4, "Add native audio input support"). The revert is clean and faithful — every hunk here is the exact inverse of a hunk in the original PR:

• /v1/audio/transcriptions handler, Codex STT forwarding (_post_audio_transcription, _resolve_audio_transcription_runtime, Codex URL/header helpers), audio MIME/format tables, and the _AUDIO_PART_TYPES rejection path in gateway/platforms/api_server.py are fully removed.
• Capability advertisement (audio_api, audio block, audio_transcriptions endpoint entry) reverted correctly — audio_api is back to False and the audio dict is gone.
• allow_auto_codex_fallback knob in hermes_cli/runtime_provider.py and the corresponding strict-mode test are removed; the vision-modality lowercasing tweak in agent/models_dev.py and the "audio" tag in run_agent.py's history summarizer are reverted to pre-merge behavior (confirmed against the original commit's diff).
• I grepped the whole repo for audio_transcriptions, AUDIO_TRANSCRIPTION, allow_auto_codex_fallback, codex_transcription, _reject_unsupported_audio, unsupported_audio_input, etc. — no leftover references outside this PR's own diff. (The hits in tests/hermes_cli/test_web_server.py for /api/audio/transcribe are an unrelated, pre-existing desktop voice-mode feature, not part of this revert.)
• Traced the history = self._conversation_history_for_session(...) relocation in _handle_session_chat_stream (gateway/platforms/api_server.py) that resulted from deleting the audio-rejection checks — it's now computed inside the _run_and_signal closure where it's used, which is correct.

One issue found

tests/gateway/test_api_server.py, _create_app() (~line 615-617) — the revert removed the /v1/audio/transcriptions route registration but also dropped this line in the same edit:

app.router.add_post("/v1/runs", adapter._handle_runs)

_handle_runs is unrelated to audio transcription and still exists/is still wired up in production (gateway/platforms/api_server.py:4840). No test in this file currently posts to /v1/runs through this test-only app builder, so nothing fails today, but this test scaffolding no longer mirrors production routing — a future test hitting /v1/runs via _create_app() would get a 404 that has nothing to do with whatever it's testing. Recommend restoring that one line.

Fix this →

Everything else looks intentional and complete — no dead code or legacy support surface left behind from the reverted feature.