upload: correlate late uploads on the call's own clock; make /upload idempotent on call_id (server-26#301, server-26#302)

The edge node now spools and retries uploads for up to 72h.

#301: when a call's started_at is more than late_upload_anchor_minutes
behind the real clock, extraction+correlation run under clock.pin(started_at)
(the replay mechanism), so recency gates and every timestamp written are on
the call's timeline. On-time uploads are unchanged. Also applies to
/calls/{id}/reprocess, which shares _run_intelligence_pipeline.

#302: an upload for a call_id whose audio is already stored returns the
stored URL without rewriting GCS or re-running Whisper/extraction/
correlation. Not atomic: two concurrent retries can both pass the check.

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

drb-c2-core/app/config.py

@@ -147,6 +147,10 @@ class Settings(BaseSettings):
     # location_proximity_km the slow path accepts). The LLM tier had no
     # distance check at all, which is how one incident drew a 37 km path.
     incident_max_spread_km: float = 2.0
+    # server-26#301: an upload arriving more than this long after the call
+    # started is correlated as of its started_at (clock pin), not now. Well
+    # above normal upload + STT lag so on-time calls are never back-dated.
+    late_upload_anchor_minutes: int = 10
     location_proximity_km: float = 0.5        # radius for location-proximity matching
     geocode_max_km: float = 40.0              # reject geocode results farther than this from the node
     incident_auto_resolve_minutes: int = 90   # auto-resolve after N minutes with no new calls (major / unknown severity)