Extraction: model in config, chatter gate, geocode cache (server-26#298, server-26#104, server-26#88)

EXTRACTION_MODEL / EMBEDDING_MODEL replace the gpt-4o-mini and
text-embedding-3-small literals. gate_chatter (classify_chatter with roll_call
narrowed to whole-transcript roll calls) skips correction, extraction,
embedding and geocoding for housekeeping, writes skip_reason=non_event_chatter
and still runs the free clearance/plate backstops; upload.py's no-scene
fallback admits it like transcript_too_short. Ships OFF (CHATTER_GATE_ENABLED
=false): a gated call has no units, so it takes the no-fit-check thin attach,
and #127 requires a clean live shadow window first.
In-process TTL cache (24h, negatives 1h, 5000 entries, LRU) in front of the
Geocoding API, keyed by address+bounds; holds coordinates only, inside the
Maps 30-day lat/lng caching limit; errors and key/quota statuses not cached.
#104: the no-scene fallback passes the call doc's severity.
#88: resolved/reassignment parsed by _llm_bool, not bool().

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

drb-c2-core/app/routers/upload.py

@@ -501,7 +501,10 @@ async def _extract_and_correlate(
     if not scenes:
         _call_doc = await fstore.doc_get("calls", call_id)
         skip_reason = (_call_doc or {}).get("skip_reason")
-        if not skip_reason or skip_reason == "transcript_too_short":
+        # non_event_chatter (server-26#298) is the same population as
+        # transcript_too_short, just longer: housekeeping that skipped the
+        # paid extraction but may still be a same-talkgroup follow-up.
+        if not skip_reason or skip_reason in ("transcript_too_short", "non_event_chatter"):
             incident_id = await _correlate_with_consensus(
                 call_id=call_id,
                 node_id=node_id,
@@ -512,6 +515,11 @@ async def _extract_and_correlate(
                 incident_type=None,
                 location=None,
                 location_coords=None,
+                # server-26#104: the call doc's own severity, as
+                # recorrelation_sweep passes it. None resolved to "routine",
+                # which turned a reprocessed not-thin call into a thin one
+                # that attaches with no fit check.
+                severity=(_call_doc or {}).get("severity"),
             )
             if incident_id:
                 incident_ids.append(incident_id)