Merge pull request 'clearance: act on drb-correlation-review of f0a88d4' (#175) from fix/clearance-review into main

Review of the first clearance fix found it pushes toward early resolve:
- parser cleared units on questions ("45-9, are you clear?"), negations
  ("not clear yet"), orders ("clear the scene", "clear to transport"),
  places/times ("Room 2 clear", "1400 hours, clear") and split "45 9" into
  unit 45. Now rejects "?", not/is/are/you, anything after the status word
  but sign-offs, place/time words; joins "45 9" -> "45-9".
- a clear from a unit never active on an incident was recorded in
  units_cleared and could pass the all-clear gate. Only units actually
  active there can clear there now.
- clearance-only calls skip the LLM tier (same as thin calls): only the
  rules engine's unit match can say which incident a 10-8 belongs to.
- replay incident view carries srcaddr/srcaddrs for the radio-ID clearance
  investigation.

Replay 09-22 10:00-12:00 ET with f0a88d4: real clears 0 -> 2 (both LLM
closure), unit clears still 0 — the parsed clears are right but those
units were never recorded as assigned (Whisper mangles unit IDs at
dispatch), which this commit does not fix.

c2-core: 465 pass.

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

drb-c2-core/app/config.py

@@ -45,7 +45,10 @@ class Settings(BaseSettings):
     # while correlation behaviour was being tuned against rules-only output.
     # Verify against https://ai.google.dev/gemini-api/docs/models before changing.
     corr_cheap_model: str = "gemini-3.6-flash"   # was gemini-2.0-flash (shut down)
-    corr_smart_model: str = "gemini-2.5-pro"     # was gemini-1.5-pro (shut down)
+    # gemini-2.5-pro was closed to new projects by 2026-09 (every tiebreak 404'd
+    # in the first replay run, server-26#170); Google lists no stable Pro model,
+    # so the smart tier is the newest stable Flash instead.
+    corr_smart_model: str = "gemini-3.8-flash"   # was gemini-2.5-pro, gemini-1.5-pro
     # Transcript correction (server-26#36). Runs inside transcription, once per
     # transcribed call above MIN_WORDS_FOR_CORRECTION, so it is priced like STT
     # rather than like the correlation tier — cheap model on purpose.

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

@@ -224,6 +224,28 @@ def _normalize_unit(unit: str) -> str:
     return key or unit.strip().lower()
 
 
+def _is_trackable_unit(unit: str) -> bool:
+    """
+    Whether a unit is concrete enough to hold an incident open until it clears.
+
+    Extraction lists everything that sounds like a unit — "Desk", "Central",
+    "Division", "sergeant", "unknown", and plate phonetics ("John Henry
+    Zebra"). None of those ever transmit a 10-8, so while they sat in
+    units_active the all-clear gate below could never pass: in the first
+    replay (server-26#170, 09-22 10:00-12:00) 0 of 19 incidents resolved on
+    a clear and every one had such a name in units_active. A real radio unit
+    ID carries a number ("45-9", "11-Adam 2", "Whitestone 1", "E-14"), so
+    only those gate resolution. The others are still kept in `units` and
+    still match for correlation.
+    """
+    if _TEN_CODE_RE.match((unit or "").strip()):
+        return False  # "10-8" read back as a unit ID is the status, not a unit
+    return any(ch.isdigit() for ch in unit or "")
+
+
+_TEN_CODE_RE = re.compile(r"^10[\s-]?\d{1,2}$")
+
+
 def _unit_keys(units: Optional[list[str]]) -> set[str]:
     """Comparison keys for a unit list, empties dropped."""
     return {k for k in (_normalize_unit(u) for u in (units or [])) if k}
@@ -1910,11 +1932,19 @@ def _apply_unit_clearance(inc: dict, cleared: list[str]) -> tuple[list[str], lis
     """
     units_active  = list(inc.get("units_active")  or [])
     units_cleared = list(inc.get("units_cleared") or [])
-    for u in cleared:
-        if u in units_active:
-            units_active.remove(u)
-        if u not in units_cleared:
+    # Compared by normalised key: the unit that cleared as "11-Adam" is the
+    # one that went active as "11 Adam", and exact equality left it active.
+    # Only a unit that was actually active here can clear here — a clear from
+    # a unit never on this incident says nothing about whether it is over, and
+    # counting it let one stray 10-8 close an incident still being worked.
+    cleared_keys = _unit_keys(cleared)
+    releasing = [u for u in units_active if _normalize_unit(u) in cleared_keys]
+    units_active = [u for u in units_active if _normalize_unit(u) not in cleared_keys]
+    known_cleared = _unit_keys(units_cleared)
+    for u in releasing:
+        if _normalize_unit(u) not in known_cleared:
             units_cleared.append(u)
+            known_cleared.add(_normalize_unit(u))
     auto_resolved = bool(units_cleared) and not units_active
     return units_active, units_cleared, auto_resolved
 
@@ -2009,9 +2039,11 @@ async def _update_incident(
     # units_active = units currently on scene; units_cleared = units back in service
     units_active  = list(inc.get("units_active")  or [])
     units_cleared = list(inc.get("units_cleared") or [])
+    tracked = _unit_keys(units_active) | _unit_keys(units_cleared)
     for u in call_units:
-        if u not in units_cleared and u not in units_active:
+        if _is_trackable_unit(u) and _normalize_unit(u) not in tracked:
             units_active.append(u)
+            tracked.add(_normalize_unit(u))
     inc_with_active_update = {**inc, "units_active": units_active, "units_cleared": units_cleared}
     units_active, units_cleared, _ = _apply_unit_clearance(inc_with_active_update, cleared_units or [])
 
@@ -2143,7 +2175,7 @@ async def _create_incident(
         "system_ids":        [system_id] if system_id else [],
         "tags":              tags + ["auto-generated"],
         "units":             call_units,
-        "units_active":      list(call_units),
+        "units_active":      [u for u in call_units if _is_trackable_unit(u)],
         "units_cleared":     [],
         "vehicles":          call_vehicles,
         "srcaddrs":          [call_srcaddr] if call_srcaddr else [],

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

@@ -247,18 +247,26 @@ async def extract_scenes(
             f"Intelligence: call {call_id} — transcript too short for extraction "
             f"({len(transcript.split())} words), skipping"
         )
+        cleared_unit = _short_clearance_unit(transcript)
         try:
             # Severity is still recorded: a five-word acknowledgement is genuinely
             # routine traffic, and downstream code treats a missing severity as
             # "not yet processed" rather than "nothing happened".
-            await fstore.doc_set("calls", call_id, {
+            updates = {
                 "skip_reason": "transcript_too_short",
                 "severity":    "routine",
                 "chatter_classifier_verdict": chatter_is_chatter,
                 "chatter_classifier_reason":  chatter_reason,
-            })
+            }
+            if cleared_unit:
+                updates["units"] = [cleared_unit]
+                updates["cleared_units"] = [cleared_unit]
+            await fstore.doc_set("calls", call_id, updates)
         except Exception:
             pass
+        if cleared_unit:
+            logger.info(f"Intelligence: call {call_id} — short clearance from {cleared_unit!r}")
+            return [_clearance_scene(transcript, cleared_unit)]
         return []
 
     try:
@@ -469,6 +477,72 @@ async def extract_scenes(
     return processed
 
 
+# "45-9, I'm clear." / "Vehicle 1, clear." / "Car 12 10-8" — a unit reporting
+# itself back in service is the one signal that ends an incident, and it is
+# almost always five words or fewer, which is exactly the population the
+# too-short skip above keeps away from GPT. In the first replay
+# (server-26#170, 09-22 10:00-12:00 ET) 25 transmissions said 10-8/clear and
+# 2 reached cleared_units. Rule-based on purpose: no model call, and only a
+# unit named BEFORE the status word counts, so "10-8, 10-8." or "CMT clear."
+# (no number) clears nobody rather than guessing.
+_CLEAR_WORD_RE = re.compile(
+    r"\b(clear|10-?8|10-?98|back in service|in service|available)\b", re.IGNORECASE
+)
+_TEN_CODE_TOKEN_RE = re.compile(r"^10-?\d{1,2}$")
+_UNIT_PREFIX_WORDS = {"unit", "car", "vehicle", "engine", "ladder", "medic", "rescue", "post", "truck", "squad"}
+# A number after one of these is a place or a time, not a radio unit.
+_NOT_UNIT_PREFIX_WORDS = {"room", "route", "rt", "exit", "pole", "apartment", "apt", "floor",
+                          "building", "hours", "hour", "block", "lane", "highway", "interstate"}
+# The status word has to END the transmission: "clear the scene", "clear to
+# transport", "available for" are orders or plans, not a unit back in service.
+_TRAILING_OK = {"10-4", "thanks", "thank", "you", "k", "over", "now", "again", "from", "headquarters", "hq",
+                "central", "dispatch"}
+
+
+def _short_clearance_unit(transcript: str) -> Optional[str]:
+    text = (transcript or "").strip()
+    if not text or "?" in text:
+        return None  # "45-9, are you clear?" asks; it doesn't report
+    m = _CLEAR_WORD_RE.search(text)
+    if not m:
+        return None
+    before = [t.strip(".,;:!") for t in text[: m.start()].split()]
+    before = [t for t in before if t]
+    after = [t.strip(".,;:!").lower() for t in text[m.end():].split()]
+    if any(t and t not in _TRAILING_OK for t in after):
+        return None
+    if any(t.lower() in {"not", "is", "are", "negative", "you"} for t in before):
+        return None  # "not clear yet", "Is 45-9 clear", "you clear"
+    for i, tok in enumerate(before[:4]):
+        if not any(ch.isdigit() for ch in tok) or _TEN_CODE_TOKEN_RE.match(tok):
+            continue
+        prev = before[i - 1].lower() if i else ""
+        if prev in _NOT_UNIT_PREFIX_WORDS:
+            return None
+        nxt = before[i + 1] if i + 1 < len(before) else ""
+        if nxt.lower() in _NOT_UNIT_PREFIX_WORDS:
+            return None  # "1400 hours, clear"
+        if nxt.isdigit():
+            tok = f"{tok}-{nxt}"  # "45 9 clear" is unit 45-9, not unit 45
+        elif nxt.isalpha() and nxt[0].isupper() and nxt.lower() not in {"i'm", "im", "we're", "copy"}:
+            tok = f"{tok} {nxt}"  # "11 Adam, clear"
+        if prev in _UNIT_PREFIX_WORDS:
+            return f"{before[i - 1]} {tok}"
+        return tok
+    return None
+
+
+def _clearance_scene(transcript: str, unit: str) -> dict:
+    """A minimal scene for a rule-parsed clearance: the unit, and nothing that
+    could make the incident-creation gate open a new incident for it."""
+    return {
+        "tags": [], "incident_type": None, "location": None, "location_coords": None,
+        "resolved": False, "severity": "routine", "vehicles": [], "units": [unit],
+        "cleared_units": [unit], "reassignment": False, "transcript": transcript,
+        "transcript_corrected": None, "segment_indices": [], "embedding": None,
+    }
+
+
 def _geo_dist_km(lat1: float, lon1: float, lat2: float, lon2: float) -> float:
     """Haversine distance in km between two lat/lon points."""
     R = 6371.0

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

@@ -275,6 +275,12 @@ async def decide(call_id: str, ctx: dict) -> Optional[dict]:
     if ctx["is_thin_call"]:
         return None  # thin calls have no transcript/units/coords to reason about
 
+    if _is_clearance_only(ctx):
+        # "45-9, I'm clear." carries one fact: which unit is done. Only the
+        # rules engine's unit match can say which incident that is; an LLM
+        # link here would apply the clear to whatever incident it picked.
+        return None
+
     if not ctx["recent"]:
         return None  # no incidents to correlate against — rules handles new-only
 
@@ -295,6 +301,14 @@ async def decide(call_id: str, ctx: dict) -> Optional[dict]:
         return None
 
 
+def _is_clearance_only(ctx: dict) -> bool:
+    units = ctx.get("call_units") or []
+    cleared = ctx.get("call_cleared") or []
+    return bool(cleared) and set(units) <= set(cleared) and not (
+        ctx.get("tags") or ctx.get("location") or ctx.get("call_vehicles") or ctx.get("incident_type")
+    )
+
+
 _dead_models: set[str] = set()
 
 

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

@@ -140,6 +140,7 @@ def _call_row(c: dict) -> dict:
         "cleared_units":  c.get("cleared_units"),
         "location":       c.get("location"),
         "skip_reason":    c.get("skip_reason"),
+        "srcaddr":        c.get("srcaddr"),
         "corr_path":      [p for p in paths if p] or ([c["corr_path"]] if c.get("corr_path") else []),
         "incident_ids":   c.get("incident_ids") or [],
     }
@@ -174,6 +175,7 @@ async def run_incidents(run_id: str, decoded: dict = Depends(require_admin_token
             "units_active":  inc.get("units_active"),
             "units_cleared": inc.get("units_cleared"),
             "talkgroup_ids": inc.get("talkgroup_ids"),
+            "srcaddrs":      inc.get("srcaddrs"),
             "calls":         rows,
         })
     orphans = sorted((r for r in by_id.values() if not r["incident_ids"]),

drb-c2-core/tests/test_clearance_tracking.py

@@ -0,0 +1,68 @@
+"""
+Dispatch→10-8 lifecycle, as measured by the first replay (server-26#170):
+0 of 19 incidents resolved on a clear although 25 transmissions said one.
+Three independent breaks, each pinned here.
+"""
+from app.internal import incident_correlator as ic
+from app.internal.intelligence import _clearance_scene, _short_clearance_unit
+
+
+def test_short_clearance_names_the_unit_that_cleared():
+    assert _short_clearance_unit("45-9, I'm clear.") == "45-9"
+    assert _short_clearance_unit("Vehicle 1, clear.") == "Vehicle 1"
+    assert _short_clearance_unit("11 Adam, clear") == "11 Adam"
+    assert _short_clearance_unit("Car 12 10-8") == "Car 12"
+    assert _short_clearance_unit("45 9 clear") == "45-9"
+    assert _short_clearance_unit("Warrant 4, clear from the jail") is None or True  # >5 words: GPT's job
+    assert _short_clearance_unit("45-9 clear, thank you") == "45-9"
+
+
+def test_short_clearance_never_guesses():
+    for t in ("10-8, 10-8.", "CMT clear.", "10-8, I'm back now. Clear.",
+              "10-8, thank you.", "Show us 10-8, post 4.", "7, Charlie Central.", "10-4.",
+              # review findings: questions, negations, orders, places, times
+              "45-9, are you clear?", "Is 45-9 clear", "45-9, not clear yet.",
+              "Engine 5 not available.", "45-9, clear the scene.", "Medic 3, clear to transport.",
+              "Room 2 clear.", "Route 9 is clear.", "1400 hours, clear.", "you clear 45-9"):
+        assert _short_clearance_unit(t) is None, t
+
+
+def test_clearance_scene_cannot_open_an_incident():
+    scene = _clearance_scene("45-9, I'm clear.", "45-9")
+    ctx = {"call_vehicles": scene["vehicles"], "coords": scene["location_coords"], "tags": scene["tags"]}
+    assert not ic.has_event_substance(ctx)
+    assert scene["severity"] == "routine" and scene["incident_type"] is None
+
+
+def test_clearance_matches_a_differently_spoken_unit():
+    inc = {"units_active": ["11 Adam", "45-9"], "units_cleared": []}
+    active, cleared, resolved = ic._apply_unit_clearance(inc, ["11-Adam"])
+    assert active == ["45-9"]
+    active, cleared, resolved = ic._apply_unit_clearance(
+        {"units_active": active, "units_cleared": cleared}, ["45 9"])
+    assert active == [] and resolved
+
+
+def test_a_unit_never_on_the_incident_cannot_close_it():
+    inc = {"units_active": ["45-9"], "units_cleared": []}
+    active, cleared, resolved = ic._apply_unit_clearance(inc, ["22-1"])
+    assert active == ["45-9"] and cleared == [] and not resolved
+    # an incident with no numbered unit ever active never resolves on a clear
+    active, cleared, resolved = ic._apply_unit_clearance({"units_active": [], "units_cleared": []}, ["22-1"])
+    assert not resolved
+
+
+def test_clearance_only_call_skips_the_llm():
+    from app.internal import llm_correlator
+    ctx = {"call_units": ["45-9"], "call_cleared": ["45-9"], "tags": [], "location": None,
+           "call_vehicles": [], "incident_type": None}
+    assert llm_correlator._is_clearance_only(ctx)
+    assert not llm_correlator._is_clearance_only({**ctx, "tags": ["mva"]})
+    assert not llm_correlator._is_clearance_only({**ctx, "call_cleared": []})
+
+
+def test_only_numbered_units_hold_an_incident_open():
+    for junk in ("Desk", "Central", "Division", "sergeant", "unknown", "John", "Zebra", "10-8", "10 4"):
+        assert not ic._is_trackable_unit(junk), junk
+    for real in ("45-9", "11-Adam", "Whitestone 1", "E-14", "Highway 3-4", "7"):
+        assert ic._is_trackable_unit(real), real

drb-c2-core/tests/test_correlator_reassignment_clearance.py

@@ -67,7 +67,10 @@ def test_clearing_a_unit_not_tracked_as_active_is_a_noop_for_active_list():
     inc = _incident(units_active=["6-7"], units_cleared=[])
     active, cleared, resolved = _apply_unit_clearance(inc, ["ghost-unit"])
     assert active == ["6-7"]
-    assert cleared == ["ghost-unit"]
+    # A unit never active on this incident is not recorded as cleared here
+    # either (server-26#170 replay review): its 10-8 says nothing about this
+    # incident, and recording it let the all-clear gate pass on a stray clear.
+    assert cleared == []
     assert resolved is False