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/tests/test_recorrelation_sweep.py

@@ -0,0 +1,97 @@
+"""
+server-26#131 — the re-correlation sweep's orphan filter checked incident_id/
+incident_ids/corr_path but had no way to tell "never processed" apart from
+"real-time pipeline (routers/upload.py _run_intelligence_pipeline) is still
+mid-flight". Racing the sweep against an in-flight real-time correlation could
+land the same call on two different incidents — the exact duplicate-link bug
+#131 found in 3 live dumps (~2% of linked calls). This pins the fix: a call
+whose intelligence_started_at marker is recent is held back from the sweep
+regardless of how orphaned it otherwise looks.
+"""
+from datetime import datetime, timezone, timedelta
+from unittest.mock import patch
+
+import pytest
+
+from app.internal import recorrelation_sweep
+
+
+def _iso(dt: datetime) -> str:
+    return dt.isoformat()
+
+
+class TestPipelineLikelyStillRunning:
+    def test_recent_marker_is_still_running(self):
+        now = datetime.now(timezone.utc)
+        call = {"intelligence_started_at": _iso(now - timedelta(minutes=1))}
+        assert recorrelation_sweep._pipeline_likely_still_running(call, now) is True
+
+    def test_old_marker_is_not_still_running(self):
+        now = datetime.now(timezone.utc)
+        call = {"intelligence_started_at": _iso(now - timedelta(minutes=30))}
+        assert recorrelation_sweep._pipeline_likely_still_running(call, now) is False
+
+    def test_marker_exactly_at_the_threshold_is_not_held_back(self):
+        # age_minutes < MIN_MINUTES_SINCE_PIPELINE_START (strict), so exactly
+        # at the threshold is old enough to release — pins the boundary so it
+        # can't drift to <= by accident and silently double the hold time.
+        now = datetime.now(timezone.utc)
+        threshold = recorrelation_sweep.MIN_MINUTES_SINCE_PIPELINE_START
+        call = {"intelligence_started_at": _iso(now - timedelta(minutes=threshold))}
+        assert recorrelation_sweep._pipeline_likely_still_running(call, now) is False
+
+    def test_no_marker_at_all_is_not_held_back(self):
+        """A pre-#131 call doc, or the marker write itself failed — absence
+        isn't evidence of an in-flight pipeline, so the sweep must still be
+        able to pick these up (that's its whole job)."""
+        now = datetime.now(timezone.utc)
+        assert recorrelation_sweep._pipeline_likely_still_running({}, now) is False
+
+    def test_unparseable_marker_is_not_held_back(self):
+        now = datetime.now(timezone.utc)
+        call = {"intelligence_started_at": "not-a-timestamp"}
+        assert recorrelation_sweep._pipeline_likely_still_running(call, now) is False
+
+
+@pytest.mark.asyncio
+async def test_sweep_pass_skips_a_call_whose_pipeline_just_started():
+    """Integration-shaped: a call that looks orphaned by every OTHER filter
+    (no incident_ids, no corr_path, no skip_reason, under the attempt budget)
+    but has a fresh intelligence_started_at must not reach correlate_call —
+    that's the race #131 found."""
+    now = datetime.now(timezone.utc)
+    racing_call = {
+        "call_id": "call-racing",
+        "started_at": _iso(now - timedelta(minutes=2)),
+        "ended_at": _iso(now - timedelta(minutes=1)),
+        "intelligence_started_at": _iso(now - timedelta(seconds=30)),
+    }
+    genuinely_orphaned_call = {
+        "call_id": "call-genuine-orphan",
+        "started_at": _iso(now - timedelta(minutes=20)),
+        "ended_at": _iso(now - timedelta(minutes=19)),
+        "intelligence_started_at": _iso(now - timedelta(minutes=19)),
+    }
+
+    async def fake_collection_where(collection, clauses):
+        assert collection == "calls"
+        return [racing_call, genuinely_orphaned_call]
+
+    correlate_calls: list[str] = []
+
+    async def fake_correlate_call(**kwargs):
+        correlate_calls.append(kwargs["call_id"])
+        return None  # no match — exercises the "not linked" branch too
+
+    doc_sets: list[tuple] = []
+
+    async def fake_doc_set(collection, doc_id, data, merge=True):
+        doc_sets.append((collection, doc_id, data))
+
+    with patch.object(recorrelation_sweep, "fstore") as mock_fstore, \
+         patch("app.internal.incident_correlator.correlate_call", fake_correlate_call):
+        mock_fstore.collection_where = fake_collection_where
+        mock_fstore.doc_set = fake_doc_set
+        await recorrelation_sweep._run_sweep_pass()
+
+    assert correlate_calls == ["call-genuine-orphan"]