Merge feat/297-cheaper-ai into feat/299-correlation-simplify

Resolve upload.py (no-scene fallback uses the new _correlate with #298's
non_event_chatter reason) and chatter_classifier docstring (gate built, off by
default). Fold intelligence._llm_bool into incident_correlator.parse_llm_bool (#88).

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

drb-c2-core/app/config.py

@@ -28,8 +28,39 @@ class Settings(BaseSettings):
 
     # OpenAI (STT + intelligence)
     openai_api_key: Optional[str] = None
+    # The OpenAI transcription model. Used when OpenAI is the STT provider AND
+    # as the fallback when another provider fails (server-26#297).
     stt_model: str = "whisper-1"  # whisper-1 | gpt-4o-mini-transcribe | gpt-4o-transcribe
 
+    # ── STT provider (server-26#297) ──────────────────────────────────────────
+    # Any OpenAI-compatible /audio/transcriptions endpoint. Groq serves
+    # whisper-large-v3-turbo at $0.04/audio-hour against whisper-1's $0.36, same
+    # API (min billed length 10s per request, prompt capped at 224 tokens --
+    # console.groq.com/docs/speech-to-text, checked 2026-10-27).
+    #   auto   -- groq when GROQ_API_KEY is set, compat when STT_BASE_URL is
+    #             set, else openai. The default, so a deploy without either
+    #             behaves exactly as before.
+    #   openai -- OpenAI only (stt_model).
+    #   groq   -- Groq (stt_compat_model), OpenAI fallback.
+    #   compat -- any other OpenAI-compatible server: stt_base_url + stt_api_key
+    #             + stt_compat_model (e.g. a self-hosted faster-whisper, #296).
+    stt_provider: str = "auto"
+    stt_base_url: Optional[str] = None   # None -> the provider's own default
+    stt_api_key: Optional[str] = None    # None -> GROQ_API_KEY / OPENAI_API_KEY
+    groq_api_key: Optional[str] = None
+    stt_compat_model: str = "whisper-large-v3-turbo"  # model on a non-OpenAI provider
+    # Retry a failed non-OpenAI request once on OpenAI stt_model. Needs
+    # OPENAI_API_KEY; without it a primary failure just means no transcript.
+    stt_fallback_to_openai: bool = True
+    # Domain vocabulary in the STT prompt (area places, local_knowledge terms,
+    # vocabulary). Applied to the non-OpenAI provider only: whisper-1 echoed
+    # vocabulary into silence before (see transcription.py), so the OpenAI
+    # path -- including the fallback -- keeps today's fixed prompt.
+    stt_vocab_prompt_enabled: bool = True
+    # Whisper keeps only the LAST 224 prompt tokens and silently drops the
+    # rest. The prompt is built against this estimated budget, with headroom.
+    stt_prompt_max_tokens: int = 200
+
     # AI request timeouts (server-26#153). Without these the OpenAI SDK waits
     # up to 600s per attempt and retries twice, and the google-genai client
     # waits forever, so a hung provider call had no upper bound at all — and
@@ -48,8 +79,31 @@ class Settings(BaseSettings):
     # google-genai does not retry unless told to, so this is the whole bound.
     gemini_timeout_seconds: float = 60.0
 
+    # Scene extraction + embeddings (server-26#298). Were literals in
+    # intelligence.py; any OpenAI chat model that supports JSON mode works.
+    extraction_model: str = "gpt-4o-mini"
+    embedding_model: str = "text-embedding-3-small"
+    # Skip correction + extraction (+ embedding, geocoding) for transcripts
+    # chatter_classifier.gate_chatter calls radio housekeeping. OFF by default:
+    # a gated call carries no units, so it is thin and takes the no-fit-check
+    # thin attach (mechanism A) instead of the fit-tested path or the #115
+    # orphan gate, and server-26#127 set "0 false positives in a live shadow
+    # window" as the bar for going live. Read chatter_classifier_verdict off a
+    # shadow window first, then set CHATTER_GATE_ENABLED=true.
+    chatter_gate_enabled: bool = False
+
     # Google Maps (geocoding)
     google_maps_api_key: Optional[str] = None
+    # In-process cache of Geocoding API answers (server-26#298). Dispatch
+    # repeats the same address on every follow-up call of a job, and the
+    # place verifier re-geocodes the same nouns. In-process rather than
+    # Firestore: a Firestore read costs about what it saves, adds latency on
+    # every miss, and would persist Maps content -- the Maps terms allow
+    # caching lat/lng for at most 30 days, and an in-memory TTL can never
+    # outlive that. Lost on restart, which only costs a few repeat lookups.
+    geocode_cache_ttl_seconds: int = 24 * 60 * 60
+    geocode_cache_negative_ttl_seconds: int = 60 * 60  # ZERO_RESULTS etc.
+    geocode_cache_max_entries: int = 5000
     # 511NY developer key (server-26#183/#198). Issued with the NYSDOT Developer
     # Access Agreement grant; the keyless legacy API is gone since the 511x cutover.
     ny511_api_key: Optional[str] = None
@@ -105,6 +159,30 @@ class Settings(BaseSettings):
     # else's PII. Both gates must be open before any snippet leaves.
     alert_transcript_snippet_enabled: bool = False
     place_verify_max_per_call: int = 3
+    # ── Alert confidence gate (server-26#300) ─────────────────────────────────
+    # Nothing in the pipeline emits a confidence, so the alerter derives one:
+    # how the rule matched (talkgroup 1.0 / extracted tag 0.8 / raw transcript
+    # text only 0.5 -- the CrimeRadar "fire alarm" -> "firearm" failure was a
+    # transcript match) scaled down when Whisper's own per-token probability
+    # (calls.stt_confidence) is below alert_stt_reference. Below the floor ->
+    # held, never pushed. High-severity matches also need the incident to
+    # have alert_corroboration_min_calls calls, or a confidence at or above
+    # alert_high_severity_min_confidence; held ones are released when a
+    # later call corroborates the incident. Starting values, not measured.
+    alert_min_confidence: float = 0.5
+    alert_high_severity_min_confidence: float = 0.85
+    alert_corroboration_min_calls: int = 2
+    # Low on purpose: exp(avg_logprob) on vocoded radio audio is often well
+    # under 0.5 for speech that is right, and Groq/OpenAI logprobs are not on
+    # a shared scale. Only clearly garbled audio should pull the score down.
+    alert_stt_reference: float = 0.3
+    alert_high_severity_terms: list[str] = [
+        "shots-fired", "shots fired", "shooting", "shot", "gun", "firearm", "weapon",
+        "stabbing", "stabbed", "officer-down", "officer down", "10-13",
+        "officer-needs-assistance", "active-shooter", "hostage", "barricade",
+        "explosion", "structure-fire", "structure fire", "working-fire",
+        "entrapment", "cardiac-arrest", "vehicle-pursuit", "pursuit",
+    ]
     # How close a candidate has to sound before it may rewrite a transcript.
     # Below this, Places Text Search will confidently hand back the nearest
     # business for any garbage string.