Alerts: confidence gate before any push, held alerts, release on corroboration (server-26#300)

No confidence field exists anywhere in scene/incident data, so the alerter
derives one: match weight (talkgroup 1.0, extracted tag 0.8, raw transcript
0.6) scaled down only when stt_confidence is under ALERT_STT_REFERENCE (0.3),
zero for not_speech. Below ALERT_MIN_CONFIDENCE, or high severity (term list,
or a major incident) without ALERT_CORROBORATION_MIN_CALLS calls or
ALERT_HIGH_SEVERITY_MIN_CONFIDENCE, the match is written to alert_holds
(no Firestore rule, admin-only GET /alerts/held) and logged instead of pushed.
A needs_corroboration hold is released when a later call brings its incident
to the call minimum; a rule pushed by that call supersedes its hold.

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

drb-c2-core/app/config.py

@@ -163,6 +163,30 @@ 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.