feat(llm): speech-to-text for patient voice input via OpenRouter (+ offline mock)

POST /api/v1/me/{token}/stt takes a 16 kHz mono WAV and only transcribes:
the patient checks the text and sends it as an ordinary message, so safety,
scrub() and the LLM gateway still apply. Audio is neither stored nor logged.
OpenRouter STT is synthetic-data only (ADR-014); SaluteSpeech for the pilot is #121.

Refs #116

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/test_api_patient_web.py

@@ -267,9 +267,9 @@ async def test_connect_care_does_not_reassign_or_create_patients(client: AsyncCl
 
 async def test_screening_capability_respects_release_gate(client: AsyncClient, monkeypatch: pytest.MonkeyPatch) -> None:
     monkeypatch.setenv("INTAKE_DEMO", "0")
-    assert (await client.get("/api/v1/patient/capabilities")).json() == {"primary_screening": False}
+    assert (await client.get("/api/v1/patient/capabilities")).json()["primary_screening"] is False
     monkeypatch.setenv("INTAKE_DEMO", "1")
-    assert (await client.get("/api/v1/patient/capabilities")).json() == {"primary_screening": True}
+    assert (await client.get("/api/v1/patient/capabilities")).json()["primary_screening"] is True
 
 
 async def test_anonymous_contact_request_reaches_doctor(client: AsyncClient, monkeypatch: pytest.MonkeyPatch) -> None:

tests/test_api_stt.py

@@ -0,0 +1,71 @@
+"""Voice input endpoint (#116): transcribes only, validates the audio, never touches the dialog."""
+
+from collections.abc import AsyncIterator
+
+import pytest
+from httpx import ASGITransport, AsyncClient
+from sqlalchemy import select
+
+from app.api import patient_web
+from app.db import session
+from app.db.bootstrap import init_db
+from app.db.clinical import Message
+from app.llm.stt import MockSTT, SttError, SttNoSpeech, load_texts
+from app.main import app
+from tests.test_llm_stt import make_wav
+
+MEM = "sqlite+aiosqlite:///:memory:"
+WAV = {"Content-Type": "audio/wav"}
+
+
+@pytest.fixture
+async def client(monkeypatch: pytest.MonkeyPatch) -> AsyncIterator[AsyncClient]:
+    monkeypatch.setattr(patient_web, "get_stt", lambda _settings: MockSTT())  # never the network in tests
+    await session.configure(identity_url=MEM, clinical_url=MEM)
+    await init_db()
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
+        yield c
+    await session.configure()
+
+
+async def _token(client: AsyncClient) -> str:
+    return (await client.post("/api/v1/auth/anonymous")).json()["token"]
+
+
+async def test_transcribes_without_writing_to_dialog(client: AsyncClient) -> None:
+    token = await _token(client)
+    resp = await client.post(f"/api/v1/me/{token}/stt", content=make_wav(2), headers=WAV)
+    assert resp.status_code == 200
+    assert resp.json() == {"text": load_texts().mock_transcript, "provider": "mock"}
+    async with session.get_clinical_session() as s:
+        assert (await s.execute(select(Message))).scalars().all() == []
+
+
+async def test_capability_and_off(client: AsyncClient, monkeypatch: pytest.MonkeyPatch) -> None:
+    assert (await client.get("/api/v1/patient/capabilities")).json()["voice_input"] is True
+    monkeypatch.setattr(patient_web, "get_stt", lambda _settings: None)
+    assert (await client.get("/api/v1/patient/capabilities")).json()["voice_input"] is False
+    token = await _token(client)
+    assert (await client.post(f"/api/v1/me/{token}/stt", content=make_wav(1), headers=WAV)).status_code == 404
+
+
+async def test_rejects_unknown_patient_bad_format_and_long_audio(client: AsyncClient) -> None:
+    assert (await client.post("/api/v1/me/nobody/stt", content=make_wav(1), headers=WAV)).status_code == 404
+    token = await _token(client)
+    url = f"/api/v1/me/{token}/stt"
+    assert (await client.post(url, content=b"OggS" + b"\0" * 200)).status_code == 415
+    assert (await client.post(url, content=make_wav(70), headers=WAV)).status_code == 413
+    assert (await client.post(url, content=make_wav(10, rate=48_000), headers=WAV)).status_code == 200
+
+
+@pytest.mark.parametrize(("error", "status"), [(SttNoSpeech("x"), 422), (SttError("x"), 503)])
+async def test_provider_errors(client: AsyncClient, monkeypatch: pytest.MonkeyPatch, error, status) -> None:
+    class Failing:
+        name = "failing"
+
+        async def transcribe(self, audio: bytes, fmt: str = "wav") -> str:
+            raise error
+
+    monkeypatch.setattr(patient_web, "get_stt", lambda _settings: Failing())
+    token = await _token(client)
+    assert (await client.post(f"/api/v1/me/{token}/stt", content=make_wav(1), headers=WAV)).status_code == status

tests/test_llm_stt.py

@@ -0,0 +1,106 @@
+"""Speech-to-text providers (#116): OpenRouter request shape, retries, silence, WAV parsing, provider choice."""
+
+import base64
+import io
+import json
+import wave
+
+import httpx
+import pytest
+
+from app.config import Settings
+from app.llm.stt import (
+    InvalidAudio,
+    MockSTT,
+    OpenRouterSTT,
+    SttError,
+    SttNoSpeech,
+    get_stt,
+    load_texts,
+    wav_seconds,
+)
+
+
+def make_wav(seconds: float, rate: int = 16_000) -> bytes:
+    buf = io.BytesIO()
+    with wave.open(buf, "wb") as w:
+        w.setnchannels(1)
+        w.setsampwidth(2)
+        w.setframerate(rate)
+        w.writeframes(b"\x00\x00" * int(rate * seconds))
+    return buf.getvalue()
+
+
+def _settings(**kw) -> Settings:
+    base = {"openrouter_api_key": "test", "stt_provider": "openrouter", "stt_timeout_s": 5.0}
+    return Settings(_env_file=None, **{**base, **kw})
+
+
+def _fake(responses: list[httpx.Response]) -> tuple[OpenRouterSTT, list[dict]]:
+    sent: list[dict] = []
+    replies = iter(responses)
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        sent.append(json.loads(request.content))
+        return next(replies)
+
+    stt = OpenRouterSTT(_settings(), client=httpx.AsyncClient(transport=httpx.MockTransport(handler)))
+    return stt, sent
+
+
+def _ok(text: str) -> httpx.Response:
+    return httpx.Response(200, json={"choices": [{"message": {"content": text}}]})
+
+
+def test_wav_seconds() -> None:
+    assert wav_seconds(make_wav(2.5)) == pytest.approx(2.5)
+    with pytest.raises(InvalidAudio):
+        wav_seconds(b"OggS" + b"\x00" * 100)
+    with pytest.raises(InvalidAudio):
+        wav_seconds(b"")
+
+
+async def test_openrouter_sends_audio_and_returns_transcript() -> None:
+    audio = make_wav(1)
+    stt, sent = _fake([_ok("«Сегодня плохо спал.»")])
+    assert await stt.transcribe(audio) == "Сегодня плохо спал."
+    body = sent[0]
+    assert body["model"] == "google/gemini-2.5-flash" and body["temperature"] == 0
+    part = body["messages"][1]["content"][1]
+    assert part["type"] == "input_audio" and part["input_audio"]["format"] == "wav"
+    assert base64.b64decode(part["input_audio"]["data"]) == audio
+
+
+async def test_silence_marker_is_no_speech() -> None:
+    stt, _ = _fake([_ok(load_texts().silence_marker.upper())])
+    with pytest.raises(SttNoSpeech):
+        await stt.transcribe(make_wav(1))
+
+
+async def test_retries_then_succeeds(monkeypatch) -> None:
+    monkeypatch.setattr("app.llm.stt.asyncio.sleep", _no_sleep)
+    stt, sent = _fake([httpx.Response(503), _ok("текст")])
+    assert await stt.transcribe(make_wav(1)) == "текст"
+    assert len(sent) == 2
+
+
+async def test_http_error_raises(monkeypatch) -> None:
+    monkeypatch.setattr("app.llm.stt.asyncio.sleep", _no_sleep)
+    stt, _ = _fake([httpx.Response(401, json={"error": {"message": "bad key"}})])
+    with pytest.raises(SttError):
+        await stt.transcribe(make_wav(1))
+
+
+async def _no_sleep(_: float) -> None:
+    return None
+
+
+async def test_mock_is_offline() -> None:
+    assert await MockSTT().transcribe(b"") == load_texts().mock_transcript
+
+
+def test_provider_choice() -> None:
+    assert isinstance(get_stt(_settings(stt_provider="auto")), OpenRouterSTT)
+    assert isinstance(get_stt(_settings(stt_provider="auto", openrouter_api_key="")), MockSTT)
+    assert get_stt(_settings(stt_provider="openrouter", openrouter_api_key="")) is None
+    assert get_stt(_settings(stt_provider="off")) is None