correlator: stop the re-correlation sweep racing an in-flight upload (#131)

server-26#131: same call_id ends up in TWO incidents' call_ids, byte-identical
extracted data, ~2% of linked calls across 3 live dumps. Root cause: the
sweep's orphan filter (incident_id/incident_ids/corr_path all absent) can't
tell 'never processed' apart from 'real-time pipeline is still mid-flight' --
a call whose STT/scene-extraction/correlation chain (routers/upload.py
_run_intelligence_pipeline) hasn't finished yet has none of those fields set,
so the sweep picks it up and correlates it independently, sometimes onto a
different incident than the real-time path lands on. Confirmed in review:
_update_incident/_create_incident append call_id to an incident's call_ids
unconditionally, with no cross-incident dedup guard -- preventing the second
correlation attempt is the only lever available at this layer.

Fix: _run_intelligence_pipeline marks intelligence_started_at on the call doc
before any slow step; the sweep holds back any call whose marker is under 15
minutes old (raised from an initial 5 -- see below), regardless of how
orphaned it otherwise looks. No marker at all (pre-#131 call doc, or the
marker write itself failed) is not held back -- absence isn't evidence of an
in-flight pipeline, and that's #131's own pre-existing population. Also
covers the /calls/{id}/reprocess path, which calls the same
_run_intelligence_pipeline.

15 min, not 5: neither the OpenAI Whisper client nor the Gemini call in
llm_correlator.py sets a request timeout (filed server-26#153), so 5 min was
a guess against an unbounded tail -- drb-correlation-review flagged this.
Raising it is free on the recovery side: a call that finished processing
(linked or genuinely orphaned) always has corr_path set and is already
excluded by the sweep's other filter, so this constant only ever delays
calls that are still actually running. DEFERRED.md row 52 (outside this
repo, Version 5C root) updated to flag its ~6 min timing figure as stale.

recorrelation_sweep.py had zero test coverage before this. New file covers
the guard function's boundary (age < threshold vs exactly-at vs old vs
missing vs unparseable) and one integration-shaped test proving a racing call
never reaches correlate_call while a genuinely-orphaned call still does.

Sandboxed pytest: 381 -> 387.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbknwttzou4s46PAykmtix

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

@@ -1461,6 +1461,21 @@ async def _apply_and_log(decision: dict, ctx: dict) -> Optional[str]:
                 "transcript":    ctx.get("scene_transcript"),
                 "incident_id":   incident_id,
                 "corr_debug":    corr_debug,
+                # server-26#139: this scene's OWN extracted incident_type/
+                # severity, as read by _call_is_substanceless's ctx at
+                # decision time — not the call doc's flat top-level field,
+                # which is last-scene-wins (server-26#96) and was the reason
+                # #138's "type" veto couldn't be told apart from cross-scene
+                # contamination without re-guessing from a live dump.
+                # NOTE: unlike incident_type, call_severity is already
+                # coerced to "routine" when extraction emitted nothing
+                # (ctx build: `severity or "routine"`) — a scene reading
+                # "routine" here doesn't distinguish "extraction said
+                # routine" from "extraction said nothing". Don't split a
+                # severity veto the way #138 splits the type veto without
+                # accounting for that.
+                "incident_type": ctx.get("incident_type"),
+                "severity":      ctx.get("call_severity"),
             }
         }
         try:

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

@@ -32,7 +32,17 @@ from app.config import settings
 # intelligence_started_at at all (pre-#131 call doc, or the marker write
 # itself failed) is NOT held back by this — absence isn't evidence of an
 # in-flight pipeline, and #131's own bug predates this field existing.
-MIN_MINUTES_SINCE_PIPELINE_START = 5
+#
+# 15, not 5: neither OpenAI's Whisper client nor Gemini's call in
+# llm_correlator.py sets a request timeout (server-26#153), so a hung call can
+# run well past a few minutes on SDK-default retries, and this constant is a
+# guess against that unbounded tail, not a measured bound. Raising it costs
+# nothing on the recovery side: a call that finished processing (linked OR
+# genuinely orphaned) always has corr_path set (_apply_and_log writes it even
+# on the orphan action), so it's already excluded by the
+# `not c.get("corr_path")` filter below and never reaches this check at all —
+# this constant only ever delays calls that are still actually running.
+MIN_MINUTES_SINCE_PIPELINE_START = 15
 
 # Standard link-only retry budget before a call is tombstoned corr_path="unlinked".
 MAX_SWEEP_ATTEMPTS = 3

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

@@ -109,6 +109,11 @@ async def debug_correlation(
             "scene_index":            scene_index,
             "transcript":             scene.get("transcript"),
             "incident_id":            scene.get("incident_id"),
+            # server-26#139: this scene's OWN incident_type/severity, as seen
+            # by _call_is_substanceless at decision time — not the call doc's
+            # flat top-level field, which is last-scene-wins (server-26#96).
+            "incident_type":          scene.get("incident_type"),
+            "severity":               scene.get("severity"),
             "corr_path":              corr_debug.get("corr_path"),
             "corr_incident_idle_min": corr_debug.get("corr_incident_idle_min"),
             "corr_distance_km":       corr_debug.get("corr_distance_km"),

drb-c2-core/tests/test_per_scene_call_doc.py

@@ -123,10 +123,66 @@ async def test_single_scene_call_still_gets_a_scenes_map_equivalent_to_flat_fiel
             "transcript": "10-4",
             "incident_id": None,
             "corr_debug": {"corr_path": "fast/thin", "corr_consensus": "rules_only"},
+            "incident_type": None,
+            "severity": None,
         }
     }
 
 
+@pytest.mark.asyncio
+async def test_scene_entry_captures_its_own_incident_type_not_a_sibling_scenes():
+    """
+    server-26#139: _call_is_substanceless's "type" veto reads ctx["incident_type"]
+    at decision time, but that value was never persisted per-scene — only the
+    last-scene-wins flat field, which #138's dump analysis couldn't
+    distinguish from cross-scene contamination. Pins _apply_and_log's write
+    side: each scene's own scenes.<n> entry carries its own incident_type/
+    severity, distinct from any other scene on the same call. Does NOT cover
+    whether the ctx handed to _call_is_substanceless is the same object that
+    reaches here — that linkage is pinned by test_consensus_gate.py and
+    test_incident_identity.py, not this file.
+    """
+    docs: dict[tuple, dict] = {}
+
+    async def fake_doc_set(collection, doc_id, data, merge=True):
+        docs.setdefault((collection, doc_id), {})
+        _merge(docs[(collection, doc_id)], data)
+
+    decision0 = {
+        "action": "orphan", "matched_incident": None, "incident_type": None,
+        "corr_debug": {"corr_path": "new", "corr_consensus": "tiebreak", "corr_gate_veto": "type"},
+    }
+    ctx0 = {
+        "call_id": "call-5", "scene_index": 0, "scene_transcript": "10-4, clear",
+        "incident_type": "traffic-stop", "call_severity": "routine",
+    }
+
+    decision1 = {
+        "action": "orphan", "matched_incident": None, "incident_type": None,
+        "corr_debug": {"corr_path": "new", "corr_consensus": "agreed"},
+    }
+    ctx1 = {
+        "call_id": "call-5", "scene_index": 1, "scene_transcript": "roll call",
+        "incident_type": None, "call_severity": "moderate",
+    }
+
+    with patch.object(incident_correlator, "fstore") as mock_fstore:
+        mock_fstore.doc_set = fake_doc_set
+        await incident_correlator._apply_and_log(decision0, ctx0)
+        await incident_correlator._apply_and_log(decision1, ctx1)
+
+    doc = docs[("calls", "call-5")]
+    scenes = doc["scenes"]
+    assert scenes["0"]["incident_type"] == "traffic-stop"
+    assert scenes["0"]["severity"] == "routine"
+    assert scenes["1"]["incident_type"] is None
+    assert scenes["1"]["severity"] == "moderate"
+    # _apply_and_log only ever flat-merges corr_debug's own keys (:1460) — a
+    # future corr_debug["incident_type"] would silently clobber
+    # intelligence.py's flat field, so this is asserted, not just commented.
+    assert "incident_type" not in doc
+
+
 @pytest.mark.asyncio
 async def test_empty_corr_debug_writes_nothing_same_as_before():
     """Preserve the pre-#96 short-circuit: no corr_debug means no write at