correlator: give the LLM tier what it needs to link, stop it defaulting to "new" (server-26#115)

The 2026-09-07 measurement window (CORRELATION_REVIEW_0907.md) showed the
consensus tiebreaker was the dominant over-split driver: it ran on 45% of
calls and resolved link/orphan disagreements as "new" ~24/25 of the time,
shattering one Mohegan Park car-alarm job into 9 incidents and opening ~7
incidents from radio checks / roll calls.

Two causes, two fixes:

1. `_inc_summary` gave the model `id|type|loc|units|tags|idle` — no title,
   no talkgroup. It literally could not see that two "car alarms, Mohegan
   Park Ave/Avenue" incidents on TG 9560 were the same. Now includes the
   incident title (the strongest same-event signal) and talkgroup.

2. `_RULES` told the model "orphan when in doubt — conservative is always
   correct". For a system that over-splits, that is backwards: a wrong link
   is cheap, a duplicate incident is the failure. Rewritten to: prefer link
   for a plausible same-talkgroup continuation (low bar), reserve "new" for a
   genuinely different event, and explicitly "orphan" non-incidents (radio
   checks, roll call, 10-8/10-98, mileage logs).

Plus `_extract_road_ids` now canonicalises street-type synonyms
(Avenue→ave, Street→st, Road→rd, ...), so "Mohegan Park Avenue" and
"Mohegan Park Ave" share a road id — that one difference was splitting the
car-alarm incident.

+tests/test_correlator_115.py. Full c2-core suite green (sandboxed venv).
Bigger levers deferred to follow-ups: the consensus escalation itself (should
a cheap-LLM "orphan" ever reach a tiebreak?), a first-class road-overlap fit
signal in _call_fits_incident, geocode coverage.

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

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

@@ -108,16 +108,31 @@ _ROAD_RE = re.compile(
 )
 
 
+# Street-type synonyms collapsed to one token so "Mohegan Park Avenue" and
+# "Mohegan Park Ave" produce the same road id (server-26#115 — that one
+# difference was splitting a car-alarm incident into two).
+_ROAD_SUFFIX_CANON = {
+    "avenue": "ave", "street": "st", "road": "rd", "drive": "dr",
+    "boulevard": "blvd", "lane": "ln", "court": "ct", "place": "pl",
+    "highway": "hwy", "parkway": "pkwy",
+}
+
+
 def _extract_road_ids(text: str) -> set[str]:
     """
     Extract normalised road/route identifiers from a location string.
     e.g. "suspect east on Route 202" → {"route 202"}
-         "at Main Street and Oak Ave" → {"main street", "oak ave"}
+         "at Main Street and Oak Ave" → {"main st", "oak ave"}
     """
-    return {
-        re.sub(r"[\s.\-]+", " ", m.group().lower()).strip()
-        for m in _ROAD_RE.finditer(text)
-    }
+    ids: set[str] = set()
+    for m in _ROAD_RE.finditer(text):
+        key = re.sub(r"[\s.\-]+", " ", m.group().lower()).strip()
+        parts = key.split()
+        if parts and parts[-1] in _ROAD_SUFFIX_CANON:
+            parts[-1] = _ROAD_SUFFIX_CANON[parts[-1]]
+            key = " ".join(parts)
+        ids.add(key)
+    return ids
 
 
 def _location_mentions_road_overlap(new_location: str, inc_mentions: list[str]) -> bool:

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

@@ -45,7 +45,18 @@ def _fmt_idle(inc: dict, now: datetime) -> str:
 
 
 def _inc_summary(inc: dict, now: datetime) -> str:
+    # server-26#115: the model was given no title and no talkgroup, so it
+    # could not tell that "car alarms, Mohegan Park Ave" and "car alarms,
+    # Mohegan Park Avenue" on the same channel were one incident — it defaulted
+    # to "new". Title is the single strongest human-readable signal for "is
+    # this the same event"; talkgroup is what makes same-channel continuation
+    # obvious.
     parts = [f"id:{inc['incident_id']}", f"type:{inc.get('type') or '?'}"]
+    tgs = inc.get("talkgroup_ids") or []
+    if tgs:
+        parts.append(f"tg:[{', '.join(str(t) for t in tgs[:3])}]")
+    if inc.get("title"):
+        parts.append(f"title:{inc['title']!r}")
     if inc.get("location"):
         parts.append(f"loc:{inc['location']}")
     units = inc.get("units") or []
@@ -88,11 +99,24 @@ def _call_block(ctx: dict) -> str:
 _SCHEMA = '{"action": "link" | "new" | "orphan", "incident_id": "<id_string or null>", "reasoning": "<one sentence>"}'
 
 _RULES = """
-Rules:
-- "link" only with clear positive evidence: same units, same geocoded location, or semantically identical scene on the same talkgroup within the last few minutes.
-- A call on a DIFFERENT talkgroup than an incident requires unit overlap or geocoded location match — topic similarity alone is not enough.
-- "new" only if the call has a clear incident_type AND describes a distinct, identifiable scene.
-- "orphan" when in doubt — conservative is always correct.
+Rules (this system OVER-SPLITS — a real incident routinely gets shattered into
+5-10 duplicates. A wrong link is cheap; a duplicate incident is the failure
+mode. Bias accordingly.):
+- Prefer "link" when the call plausibly continues a recent incident ON THE SAME
+  TALKGROUP: same or overlapping units, the same or an adjacent location (treat
+  "Ave"/"Avenue", "St"/"Street", "Rd"/"Road" as identical; a house number plus
+  the same street is the same place), the same subject/vehicle/case number, or a
+  follow-up beat ("units clearing", "negative contact", "tow en route", "event
+  number 214-201", a status update) to an incident that is only a few minutes
+  idle. The bar for "link" on the same talkgroup is LOW.
+- Reserve "new" for a call that clearly describes a DIFFERENT event from every
+  recent incident — a different place, different units, and a different subject,
+  not merely a different transmission about the same job.
+- "orphan" a call that is not an incident at all: radio checks, roll call,
+  a unit marking on/off duty or 10-8/10-98, mileage/log entries, a bare
+  acknowledgement. Do not open a "new" incident for these.
+- A call on a DIFFERENT talkgroup than an incident still requires unit overlap
+  or a geocoded/location match — topic similarity alone is not enough there.
 - Do NOT link just because both calls involve police or both mention a road.
 """