Transcription: config-driven OpenAI-compatible STT with OpenAI fallback (server-26#297, server-26#83)

STT_PROVIDER=auto picks Groq whisper-large-v3-turbo only when GROQ_API_KEY is
set (else OpenAI, unchanged); compat covers any OpenAI-compatible server via
STT_BASE_URL/STT_API_KEY. A failed primary falls back to OpenAI STT_MODEL on
the same downloaded audio; the fallback reports to its own ai_health tier so
a dead primary stays red. The non-OpenAI provider gets a token-budgeted
glossary prompt (area places, local_knowledge, vocabulary; never ten-codes),
guarded by a narrow prompt-echo check. A per-30s-window zlib compression
check (Whisper's 2.4) adds a noise gate that survives correction being off
(#83, partial). Call docs record stt_provider/model/audio_seconds/confidence.
The sweep hold-back counts the fallback attempt. gate_chatter added to
chatter_classifier (used by transcription to skip correction when
CHATTER_GATE_ENABLED). GROQ_API_KEY templated with default ''.

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

drb-c2-core/app/internal/chatter_classifier.py

@@ -14,13 +14,17 @@ This module is that classifier. It is a PURE function of the transcript text —
 no Firestore, no LLM call, no side effects — so it is cheap to run on every
 transcript and cheap to test against real dumps offline.
 
-SHADOW MODE ONLY. As of this module's introduction, nothing skips scene
-extraction based on this verdict. `intelligence.extract_scenes` calls
-`classify_chatter` purely to record the verdict on the call doc
-(`chatter_classifier_verdict` / `chatter_classifier_reason`) so it becomes
-observable in the next `/admin` correlation-debug dump, exactly like
-`corr_gate_veto` (server-26#115 / PR #126). See the TODO at that call site for
-what has to be true before this flips live.
+LIVE AS A COST GATE SINCE server-26#298, through `gate_chatter`, not
+`classify_chatter`. `classify_chatter` is still the shadow verdict recorded on
+every call doc (`chatter_classifier_verdict` / `chatter_classifier_reason`),
+unchanged, so the `/admin` tallies keep measuring the same thing. The gate is
+stricter on the one shape the #127 review flagged as a live false-positive
+risk: `roll_call` matches "roll call" anywhere, so "Delayed for roll call, then
+respond to a structure fire at 12 Main Street" is a roll_call verdict. The gate
+only skips a roll call whose remaining words are callsigns and roll-call
+phrasing. Gated transcripts skip transcript correction and scene extraction
+(and with it embedding and geocoding); the free rule paths (clearance parse,
+self-initiated/plate backstop) still run.
 
 Precision over recall, deliberately. A false positive here — flagging a REAL
 event as chatter — would, once live, silently mean that event never gets a
@@ -133,3 +137,36 @@ def classify_chatter(transcript: Optional[str]) -> tuple[bool, Optional[str]]:
         return False, None
 
     return True, "bare_acknowledgement"
+
+
+# Words a roll call is made of, beyond the filler/designator/phonetic sets.
+_ROLL_CALL_WORDS = frozenset({
+    "roll", "call", "all", "cars", "car", "patrol", "patrols", "troop", "staff",
+    "concludes", "time", "came", "because", "volunteer", "nyc", "kl", "henry",
+    "here", "present", "s", "m", "it",
+})
+# A roll call may name a place or two ("Post 4, Ossining"); more than this many
+# leftover words means something else is being said in the same transmission.
+_ROLL_CALL_MAX_LEFTOVER = 2
+
+
+def gate_chatter(transcript: Optional[str]) -> tuple[bool, Optional[str]]:
+    """
+    Whether to skip paid AI for this transcript (server-26#298).
+
+    `classify_chatter`'s verdict, with roll_call narrowed: the transcript must
+    be roll-call content end to end, not a roll call mentioned on the way to
+    something real. Precision over recall still — a false positive here means
+    a real event never gets a scene.
+    """
+    is_chatter, reason = classify_chatter(transcript)
+    if not is_chatter:
+        return False, None
+    if reason == "roll_call":
+        leftover = [
+            t for t in _tokenize(transcript or "")
+            if not _is_filler_token(t) and t not in _ROLL_CALL_WORDS
+        ]
+        if len(leftover) > _ROLL_CALL_MAX_LEFTOVER:
+            return False, None
+    return True, reason