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/recorrelation_sweep.py

@@ -20,6 +20,30 @@ from app.internal.logger import logger
 from app.internal import firestore as fstore
 from app.config import settings
 
+# server-26#131: minimum time since the real-time pipeline (routers/upload.py
+# _run_intelligence_pipeline) marked intelligence_started_at before the sweep
+# will touch a call, even though it already looks orphaned. STT + scene
+# extraction + correlation is a multi-second-to-low-minutes chain (Whisper,
+# then a Gemini call per scene); without this buffer the sweep could pick up
+# a call mid-pipeline — no incident_id/corr_path written yet — and correlate
+# it a second time, independently, sometimes onto a different incident than
+# the real-time path lands on. That race is what #131 found (same call in
+# two incidents' call_ids, ~2% of linked calls). A call with no
+# 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.
+#
+# 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
 # server-26#115 — a call the consensus LLM-orphan gate parked (llm=orphan vs
@@ -52,8 +76,22 @@ async def recorrelation_loop() -> None:
             logger.error(f"Re-correlation sweep failed: {e}")
 
 
+def _pipeline_likely_still_running(call: dict, now: datetime) -> bool:
+    """server-26#131 — True when the real-time pipeline marked
+    intelligence_started_at recently enough that it's probably still mid-flight
+    (STT / scene extraction / correlation), so the sweep should not race it.
+    No marker at all (older call doc, or the marker write itself failed)
+    returns False — absence isn't evidence of an in-flight pipeline."""
+    started = _parse_dt(call.get("intelligence_started_at"))
+    if not started:
+        return False
+    age_minutes = (now - started).total_seconds() / 60
+    return age_minutes < MIN_MINUTES_SINCE_PIPELINE_START
+
+
 async def _run_sweep_pass() -> None:
-    cutoff = datetime.now(timezone.utc) - timedelta(minutes=settings.recorrelation_scan_minutes)
+    now = datetime.now(timezone.utc)
+    cutoff = now - timedelta(minutes=settings.recorrelation_scan_minutes)
 
     # Server-side range query: only calls that ended within the scan window.
     # Filter incident_id=null client-side (Firestore can't query for missing fields).
@@ -77,6 +115,7 @@ async def _run_sweep_pass() -> None:
         # a second route into the over-merge the thin fix above addresses.
         and not c.get("skip_reason")
         and c.get("corr_sweep_count", 0) < _max_sweep_attempts(c)
+        and not _pipeline_likely_still_running(c, now)
     ]
 
     if not orphans: