intelligence: run the chatter classifier before the too-short skip, not after (#127)

82% of the classifier's backtest flags were <=5-word transcripts that already exit at skip_reason=transcript_too_short before the classifier ever ran, so shadow mode was on track to observe roughly a fifth of the real catch rate. Compute the verdict once, ahead of that check, and fold it into whichever doc_set already runs (no extra Firestore write). Also add a chatter_classifier_flagged/reason tally spanning both linked calls AND orphans in admin.py's summary block -- the target population is non-events, which land as orphans or single-call incidents, so linked alone undercounts it the same way corr_gate_veto would have without the #126 fix.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Claude-Session: https://claude.ai/code/session_01Tbknwttzou4s46PAykmtix

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

@@ -200,6 +200,28 @@ async def extract_scenes(
             pass
         return []
 
+    # server-26#127 — SHADOW MODE ONLY. Computes whether this transcript looks
+    # like non-event radio housekeeping (roll call, bare 10-4/10-8/98
+    # acknowledgements, unit check-ins) and records the verdict on the call
+    # doc, but does NOT skip extraction anywhere below — every path runs
+    # exactly as it did before this landed. Deliberately ahead of the ≤5-word
+    # skip: most bare acknowledgements ARE ≤5 words, and the first pass of
+    # this feature put the classifier after that return, so it never saw the
+    # bulk of its own target population — a review backtest against three
+    # live dumps found 82% of what it would have flagged already exits above
+    # as transcript_too_short, meaning a shadow-mode window would have shown
+    # roughly a fifth of the real catch rate. Computing it once, here, and
+    # folding the result into whichever skip/continue path runs below fixes
+    # that without adding a second Firestore write.
+    # TODO(server-26#127): flip this from shadow to live (skip extraction and
+    # write skip_reason="non_event_chatter" instead of just recording the
+    # verdict) once a live shadow-mode window confirms 0 false positives on
+    # real production traffic — pay particular attention to whole-transcript
+    # vs contains-anywhere matching for "roll call" and to digit-hyphen street
+    # addresses (e.g. "72-Holland"), both flagged as classifier risks that the
+    # dump backtest could not surface on its own.
+    chatter_is_chatter, chatter_reason = classify_chatter(transcript)
+
     # Transcripts with ≤5 words carry no extractable intelligence — GPT hallucinates
     # units and tags from thin context (e.g. "Main Lot", "10-4", "David").
     if len(transcript.split()) <= 5:
@@ -214,23 +236,13 @@ async def extract_scenes(
             await fstore.doc_set("calls", call_id, {
                 "skip_reason": "transcript_too_short",
                 "severity":    "routine",
+                "chatter_classifier_verdict": chatter_is_chatter,
+                "chatter_classifier_reason":  chatter_reason,
             })
         except Exception:
             pass
         return []
 
-    # server-26#127 — SHADOW MODE ONLY. Computes whether this transcript looks
-    # like non-event radio housekeeping (roll call, bare 10-4/10-8/98
-    # acknowledgements, unit check-ins) and records the verdict on the call
-    # doc, but does NOT skip extraction — the scene-extraction call below runs
-    # exactly as it does today regardless of what this says. This is step one
-    # of getting live production data on the classifier's false-positive rate
-    # before trusting it with anything real; see CORRELATION_REVIEW_0912.md.
-    # TODO(server-26#127): flip this from shadow to live (skip extraction and
-    # write skip_reason="non_event_chatter" instead of just recording the
-    # verdict) once a live shadow-mode window confirms 0 false positives on
-    # real production traffic.
-    chatter_is_chatter, chatter_reason = classify_chatter(transcript)
     try:
         await fstore.doc_set("calls", call_id, {
             "chatter_classifier_verdict": chatter_is_chatter,

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

@@ -309,6 +309,17 @@ async def debug_correlation(
         # server-26#115 — this IS the number the escape-hatch fix exists to
         # produce: why each llm=orphan/rules=new call escaped the gate.
         "corr_gate_veto":         _tally(c.get("corr_gate_veto") for c in linked),
+        # server-26#127 — shadow-mode chatter classifier. The target
+        # population is non-events, which land as orphans or single-call
+        # incidents, NOT as a slice of every linked call -- tally `orphans`
+        # too or this undercounts the exact thing the feature measures.
+        "chatter_classifier_flagged": sum(
+            1 for c in (linked + orphans) if c.get("chatter_classifier_verdict")
+        ),
+        "chatter_classifier_reason": _tally(
+            c.get("chatter_classifier_reason") for c in (linked + orphans)
+            if c.get("chatter_classifier_verdict")
+        ),
         # STT coverage: correlation quality is capped by this, so it belongs in
         # the same view rather than a separate investigation.
         "linked_calls_with_transcript": with_transcript,