intelligence: don't reject a geocode just because it is far from the node (#159)

_geocode_location's node-distance fallback (geocode_max_km, 40km) is a
proxy for "is this plausible" that only makes sense when the node's own
position is the best guess we have at the area. It doesn't hold for a
citywide/patched feed: node-002 sits in Westchester but relays "New York
City - NYPD Citywide 2 Patch", ~56km from the addresses on it. Every real,
correctly-geocoded address on that talkgroup was rejected by this check,
every time -- location_coords stayed permanently null for the whole
system, which killed the location_proximity correlation signal and let
duplicate incidents form for the same event reported at two nearby
addresses two minutes apart ("Shots Fired at Jackson Avenue" /
"Shots Fired at 1108 Jackson Avenue", 2026-09-20 ~23:00 UTC, merged by
hand via the Archive page's attach/detach while this fix went in).

trust_named_region skips the node-distance rejection exactly when the
query already carries a place name that isn't the node's own position --
operator-set area_context, or a municipality parsed from the talkgroup's
own name. The anchor path is untouched; an anchor's own radius is always
authoritative when one has been resolved.

Also fixes a compounding defect found while verifying: the query was
grafting the node's own COUNTY onto an already-self-named region
("...New York City..., Westchester, New York"), which is self-contradictory
and could degrade the geocode independent of the distance check. The
node's county is now used only when nothing else names the place; state
stays in both branches since it's coarse enough to be correct either way.
Extracted the query-assembly logic into a pure, unit-tested helper
(_location_query_parts) rather than testing it only through the full
extraction pipeline.

Filed #160 as a follow-up: place_verifier.py's verify() has the identical
no-anchor gap for transcript place-name correction, not fixed here.

Verified: 422 pass, 0 fail (9 new tests).

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

.gitea/workflows/deploy.yml

@@ -115,27 +115,14 @@ jobs:
           # Update compose files + mosquitto config
           git pull origin main
 
-          # server-26#51: Firestore rules + composite indexes had no deploy
+          # server-26#51: Firestore rules/indexes deploy used to be attempted
-          # path and regressed silently after every fix (the alert_events and
+          # HERE, over SSH, gated on the VM having firebase-tools installed.
-          # calls(org_id,started_at) indexes among them). The VM runs as the
+          # It never did (no node on the VM), so this silently warned and
-          # project service account, so firebase-tools authenticates via ADC
+          # skipped on every deploy for weeks -- PR #124 even auto-closed
-          # with no key file, and infra/firestore/firebase.json pins database
+          # #13/#51 as if it were fixed. Moved to a standalone
-          # c2-server. Indexes go on additively -- no --force -- so a stray
+          # deploy-firestore-rules job below that runs on the Gitea runner
-          # edit to firestore.indexes.json can never delete a live index;
+          # itself (which always has node), so it no longer depends on
-          # rules are a full replace, which is the intent. --non-interactive
+          # anything being pre-installed on this VM.
-          # means the FIRST run after a drift still needs a one-time manual
-          # `firebase deploy` on the VM to clear pending deletions (it aborts
-          # rather than guess). A failure here warns but does NOT fail the
-          # deploy: a transient Firebase API error must not roll back a good
-          # app build.
-          if command -v firebase >/dev/null 2>&1; then
-            ( cd /opt/drb/infra/firestore \
-              && firebase deploy --only firestore:rules,firestore:indexes \
-                   --project ${{ secrets.FIREBASE_PROJECT_ID }} --non-interactive ) \
-              || echo "WARNING: firestore deploy failed (server-26#51) -- rules/indexes may be stale"
-          else
-            echo "WARNING: firebase CLI not on the VM -- skipped firestore deploy (server-26#51); install once with: npm i -g firebase-tools"
-          fi
 
           # server-26#65: capture what is actually live BEFORE switching, so
           # a bad deploy has something concrete to fall back to. This reads
@@ -291,9 +278,48 @@ jobs:
           echo "status=success" >> "$GITHUB_OUTPUT"
           echo "rolled_back_to=$PREV_TAG" >> "$GITHUB_OUTPUT"
 
+  deploy-firestore-rules:
+    name: Deploy Firestore rules & indexes
+    needs: build
+    runs-on: ubuntu-latest
+    # Deliberately independent of the `deploy` job (app containers) and its
+    # health-check/rollback chain above: a rules/indexes deploy failure has
+    # nothing to roll back (there is no previous "build" of a ruleset to
+    # revert to via this pipeline) and must never be conflated with an app
+    # deploy failure by triggering that job's rollback logic. This job
+    # failing is its own, separate red run -- picked up by notify-failure
+    # below -- not a signal to touch the running containers.
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Deploy firestore rules and indexes
+        env:
+          FIREBASE_TOKEN: ${{ secrets.FIREBASE_TOKEN }}
+        run: |
+          set -e
+          # server-26#51: this used to run over SSH on the deploy VM, gated
+          # on the VM having firebase-tools installed. It never did, so it
+          # silently warned-and-skipped on every single deploy for weeks.
+          # Running it here instead means the only prerequisite is a secret
+          # -- FIREBASE_TOKEN, from `firebase login:ci` -- rather than
+          # something installed by hand on a machine this pipeline doesn't
+          # otherwise touch. A missing token now fails this job LOUDLY
+          # (picked up by notify-failure) instead of a buried warning line
+          # nobody reads in the app deploy's logs.
+          if [ -z "$FIREBASE_TOKEN" ]; then
+            echo "FIREBASE_TOKEN secret is not set -- cannot deploy Firestore rules/indexes." >&2
+            echo "Generate one with 'firebase login:ci' and add it as a Gitea Actions secret." >&2
+            exit 1
+          fi
+          npm install -g firebase-tools
+          cd infra/firestore
+          firebase deploy --only firestore:rules,firestore:indexes \
+            --project ${{ secrets.FIREBASE_PROJECT_ID }} \
+            --token "$FIREBASE_TOKEN" --non-interactive
+
   notify-failure:
     name: Report a failed deploy
-    needs: [build, deploy]
+    needs: [build, deploy, deploy-firestore-rules]
     if: failure()
     runs-on: ubuntu-latest
 
@@ -310,6 +336,8 @@ jobs:
           SHA: ${{ gitea.sha }}
           ROLLBACK_STATUS: ${{ needs.deploy.outputs.rollback_status }}
           ROLLBACK_SHA: ${{ needs.deploy.outputs.rollback_sha }}
+          DEPLOY_RESULT: ${{ needs.deploy.result }}
+          RULES_RESULT: ${{ needs.deploy-firestore-rules.result }}
         run: |
           if [ -z "$WEBHOOK" ]; then
             echo "DEPLOY_ALERT_WEBHOOK is not set - skipping notification."
@@ -321,6 +349,16 @@ jobs:
           run_url = os.environ["RUN_URL"]
           status = os.environ.get("ROLLBACK_STATUS", "")
           rollback_sha = os.environ.get("ROLLBACK_SHA", "")
+          deploy_result = os.environ.get("DEPLOY_RESULT", "")
+          rules_result = os.environ.get("RULES_RESULT", "")
+
+          # deploy-firestore-rules runs independent of the app deploy/rollback
+          # chain (see its own job comment), so its failure needs its own
+          # branch here -- otherwise this fell through to the generic "Build
+          # failed before any deploy was attempted" text even when the app
+          # deployed fine and only the Firestore rules/indexes push failed.
+          if deploy_result != "failure" and rules_result == "failure":
+              detail = "App deploy succeeded; Firestore rules/indexes deploy FAILED (server-26#51). Rules may be stale — check FIREBASE_TOKEN and the job log."
 
           # server-26#65: the old text here unconditionally claimed
           # "production is still running the previous build" -- true only
@@ -329,7 +367,7 @@ jobs:
           # class of bug the correlator instrumentation exists to catch), or
           # once the deploy job's own rollback path has run. Say what
           # actually happened instead.
-          if status == "success":
+          elif status == "success":
               detail = "Automatic rollback to `%s` succeeded. Production is back on the previous good build." % rollback_sha[:8]
           elif status == "failed":
               detail = ("Automatic rollback to `%s` FAILED. Production state is UNKNOWN -- "

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

@@ -875,11 +875,17 @@ async def _build_context(
     is_thin_call   = _is_thin_call(
         call_units, call_vehicles, coords, tags, location, call_severity, reassignment
     )
+    # server-26#158: the P25 source radio ID. Captured on every call by the
+    # edge node's metadata_watcher.py independent of transcript content, so
+    # it survives even when transcript_too_short skips GPT extraction
+    # entirely and leaves call_units empty — exactly the population the
+    # thin-call path below has no other identity signal for.
+    call_srcaddr   = call_doc.get("srcaddr")
 
     return {
         "call_id": call_id, "org_id": org_id, "all_active": all_active, "recent": recent,
         "call_doc": call_doc, "call_embedding": call_embedding,
-        "scene_transcript": scene_transcript,
+        "scene_transcript": scene_transcript, "call_srcaddr": call_srcaddr,
         "call_units": call_units, "call_vehicles": call_vehicles,
         "call_cleared": call_cleared, "call_severity": call_severity,
         "coords": coords, "is_thin_call": is_thin_call, "now": now,
@@ -933,6 +939,7 @@ def _run_decision(ctx: dict) -> dict:
     call_severity  = ctx["call_severity"]
     coords         = ctx["coords"]
     is_thin_call   = ctx["is_thin_call"]
+    call_srcaddr   = ctx.get("call_srcaddr")
     system_id      = ctx["system_id"]
     talkgroup_id   = ctx["talkgroup_id"]
     talkgroup_name = ctx["talkgroup_name"]
@@ -1008,32 +1015,52 @@ def _run_decision(ctx: dict) -> dict:
             # incident idle up to tg_fast_path_idle_minutes (90) with no
             # single-candidate requirement and no fit test of any kind.  Four
             # hours is not a bound, and neither is ninety minutes.
+            # server-26#158: identity beats guesswork. A thin call has no
+            # extracted units (GPT never ran), but it still carries the P25
+            # radio ID that transmitted it — stronger, cheaper evidence than
+            # "most recently active" and immune to the exact failure this
+            # path exists to guard against: two incidents both live on one
+            # busy dispatch channel. If the radio that sent this call already
+            # has calls on one of the TG-matched incidents, that IS the
+            # thread, regardless of which incident is more recently updated
+            # or how many candidates are in the window.
+            srcaddr_matches = [
+                inc for inc in tg_recent
+                if call_srcaddr and call_srcaddr in (inc.get("srcaddrs") or [])
+            ]
             THIN_CONVERSATIONAL_SECS = 30
             thin_window_min = settings.tg_dispatch_thin_idle_minutes
-            very_recent = [
+            if srcaddr_matches:
-                inc for inc in tg_recent
+                thin_pool = [max(srcaddr_matches, key=lambda inc: inc.get("updated_at", ""))]
-                if _idle_gate_minutes(inc, now) * 60 <= THIN_CONVERSATIONAL_SECS
-            ]
-            if very_recent:
-                # Tier 1: direct conversational reply — most recent wins.
-                thin_pool = [max(very_recent, key=lambda inc: inc.get("updated_at", ""))]
                 logger.info(
-                    f"Correlator fast-path thin (tier-1, ≤{THIN_CONVERSATIONAL_SECS}s): "
+                    f"Correlator fast-path thin (srcaddr match): radio {call_srcaddr} "
-                    f"using most-recent of {len(very_recent)} candidate(s) for call {call_id}"
+                    f"already on {len(srcaddr_matches)} candidate(s) for call {call_id}"
                 )
             else:
-                # Tier 2: less certain — require a single candidate inside the
+                very_recent = [
-                # channel's thin window.
-                thin_pool = [
                     inc for inc in tg_recent
-                    if _idle_gate_minutes(inc, now) <= thin_window_min
+                    if _idle_gate_minutes(inc, now) * 60 <= THIN_CONVERSATIONAL_SECS
                 ]
-                if len(thin_pool) > 1:
+                if very_recent:
+                    # Tier 1: direct conversational reply — most recent wins.
+                    thin_pool = [max(very_recent, key=lambda inc: inc.get("updated_at", ""))]
                     logger.info(
-                        f"Correlator fast-path thin (tier-2): {len(thin_pool)} active incidents "
+                        f"Correlator fast-path thin (tier-1, ≤{THIN_CONVERSATIONAL_SECS}s): "
-                        f"— ambiguous, skipping thin call {call_id}"
+                        f"using most-recent of {len(very_recent)} candidate(s) for call {call_id}"
                     )
-                    thin_pool = []
+                else:
+                    # Tier 2: less certain — require a single candidate inside the
+                    # channel's thin window.
+                    thin_pool = [
+                        inc for inc in tg_recent
+                        if _idle_gate_minutes(inc, now) <= thin_window_min
+                    ]
+                    if len(thin_pool) > 1:
+                        logger.info(
+                            f"Correlator fast-path thin (tier-2): {len(thin_pool)} active incidents "
+                            f"— ambiguous, skipping thin call {call_id}"
+                        )
+                        thin_pool = []
 
             if not thin_pool:
                 logger.info(
@@ -1049,8 +1076,10 @@ def _run_decision(ctx: dict) -> dict:
                     # no fit signal, so the admin debug view's "fit_signal
                     # distribution" panel read empty on 95% of calls and looked
                     # broken. Name what actually decided it: recency on this
-                    # talkgroup, with no content to check a fit against.
+                    # talkgroup, with no content to check a fit against — or,
-                    "corr_fit_signal": "thin_recency",
+                    # when the same radio ID already touched a candidate
+                    # (server-26#158), that identity match instead of a guess.
+                    "corr_fit_signal": "thin_srcaddr_match" if srcaddr_matches else "thin_recency",
                     "corr_candidates": len(thin_pool),
                 }
                 logger.info(
@@ -1511,6 +1540,7 @@ async def _apply_decision(decision: dict, ctx: dict) -> Optional[str]:
     call_cleared    = ctx["call_cleared"]
     coords          = ctx["coords"]
     now             = ctx["now"]
+    call_srcaddr    = ctx.get("call_srcaddr")
     incident_type   = decision["incident_type"]
 
     if action == "link":
@@ -1522,7 +1552,7 @@ async def _apply_decision(decision: dict, ctx: dict) -> Optional[str]:
             location, location_coords, call_units, call_vehicles, call_embedding, now,
             talkgroup_name=talkgroup_name, incident_type=incident_type,
             cleared_units=call_cleared, refresh_activity=not thin_link,
-            call_severity=call_severity,
+            call_severity=call_severity, call_srcaddr=call_srcaddr,
         )
         return matched_incident["incident_id"]
 
@@ -1554,6 +1584,7 @@ async def _apply_decision(decision: dict, ctx: dict) -> Optional[str]:
             call_id, org_id, incident_type, talkgroup_id, talkgroup_name, system_id,
             tags, location, location_coords,
             call_units, call_vehicles, call_embedding, call_severity, now,
+            call_srcaddr=call_srcaddr,
         )
 
         if existing_master_id:
@@ -1599,6 +1630,7 @@ async def _apply_decision(decision: dict, ctx: dict) -> Optional[str]:
             call_id, org_id, incident_type, talkgroup_id, talkgroup_name, system_id,
             tags, location, location_coords,
             call_units, call_vehicles, call_embedding, call_severity, now,
+            call_srcaddr=call_srcaddr,
         )
         decision["corr_debug"]["corr_path"] = "new"
 
@@ -1945,6 +1977,7 @@ async def _update_incident(
     cleared_units: Optional[list[str]] = None,
     refresh_activity: bool = True,
     call_severity: Optional[str] = None,
+    call_srcaddr: Optional[str] = None,
 ) -> None:
     incident_id = inc["incident_id"]
 
@@ -1963,6 +1996,12 @@ async def _update_incident(
     merged_tags     = list(dict.fromkeys((inc.get("tags")     or []) + tags))
     merged_units    = list(dict.fromkeys((inc.get("units")    or []) + call_units))
     merged_vehicles = list(dict.fromkeys((inc.get("vehicles") or []) + call_vehicles))
+    # server-26#158: accumulate every radio ID that has transmitted on this
+    # incident, so a later thin call from the same radio can identity-match
+    # instead of guessing off recency alone.
+    merged_srcaddrs = list(dict.fromkeys(
+        (inc.get("srcaddrs") or []) + ([call_srcaddr] if call_srcaddr else [])
+    ))
 
     # Unit activity tracking: units_active / units_cleared
     # units_active = units currently on scene; units_cleared = units back in service
@@ -1993,6 +2032,7 @@ async def _update_incident(
         "tags":              merged_tags,
         "units":             merged_units,
         "vehicles":          merged_vehicles,
+        "srcaddrs":          merged_srcaddrs,
         "units_active":      units_active,
         "units_cleared":     units_cleared,
         "location_mentions": location_mentions,
@@ -2059,6 +2099,7 @@ async def _create_incident(
     call_embedding: Optional[list],
     call_severity: str,
     now: datetime,
+    call_srcaddr: Optional[str] = None,
 ) -> str:
     incident_id = str(uuid.uuid4())
     tg_label = (
@@ -2102,6 +2143,7 @@ async def _create_incident(
         "units_active":      list(call_units),
         "units_cleared":     [],
         "vehicles":          call_vehicles,
+        "srcaddrs":          [call_srcaddr] if call_srcaddr else [],
         "severity":          call_severity,
         "summary":           None,
         "summary_stale":     True,

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

@@ -368,18 +368,20 @@ async def extract_scenes(
         # the country.
         location_coords: Optional[dict] = None
         if location:
-            parts = [location]
+            node_state, node_county = "", ""
-            if tg_area.get("municipality") or tg_area.get("county") or tg_area.get("state"):
+            if not area_context.has_place(tg_area) and node_id and node_lat is not None and node_lon is not None:
-                parts += [tg_area[f] for f in area_context.PLACE_FIELDS if tg_area.get(f)]
+                # Only worth the (cached-after-first-call) reverse-geocode
-            elif node_lat is not None and node_lon is not None:
+                # when nothing better already describes this talkgroup.
-                muni   = _municipality_from_tg(talkgroup_name)
+                node_state  = await _get_node_state(node_id, node_lat, node_lon)
-                state  = await _get_node_state(node_id or "", node_lat, node_lon) if node_id else ""
+                node_county = _node_county_cache.get(node_id) or ""
-                county = _node_county_cache.get(node_id or "") if node_id else ""
+            parts, tg_named_region = _location_query_parts(
-                parts += [p for p in (muni, county, state) if p]
+                location, tg_area, talkgroup_name, node_state, node_county,
+            )
             query = ", ".join(parts)
             if tg_anchor or (node_lat is not None and node_lon is not None):
                 location_coords = await _geocode_location(
-                    query, node_lat, node_lon, anchor=tg_anchor
+                    query, node_lat, node_lon, anchor=tg_anchor,
+                    trust_named_region=tg_named_region,
                 )
 
         # Embed this scene's content
@@ -514,6 +516,7 @@ async def _geocode_location(
     node_lat: Optional[float] = None,
     node_lon: Optional[float] = None,
     anchor: Optional[dict] = None,
+    trust_named_region: bool = False,
 ) -> Optional[dict]:
     """
     Geocode using Google Maps Geocoding API, biased toward the channel's area.
@@ -529,6 +532,26 @@ async def _geocode_location(
     talkgroup has a resolved anchor, that is the reference and its own radius is
     the bound. Distance-from-node stays only as the fallback for a system nobody
     has described yet — it was always a stand-in for this.
+
+    server-26#159: "a system nobody has described yet" turned out to include
+    systems that describe themselves — "New York City - NYPD Citywide 2 Patch"
+    names its own coverage area right in the talkgroup name, parsed into the
+    query by `_municipality_from_tg`, but a large aggregated/patched feed like
+    this is routinely received 40-70km from an antenna that happens to sit
+    wherever the node owner lives. Real, correctly-geocoded addresses on that
+    feed were being rejected by the node-distance check every single time —
+    location_coords stayed permanently null for the whole system, which killed
+    the location_proximity correlation signal and let duplicate incidents form
+    for the same event reported at two nearby addresses two minutes apart.
+
+    `trust_named_region` is True exactly when the query already carries a place
+    name that isn't the node's own position — operator-set area_context, or a
+    municipality parsed from the talkgroup's own name. In that case a distant
+    node is not evidence of a bad geocode, so the node-distance check is
+    skipped and precision is judged by `location_type` alone (still required
+    to be ROOFTOP/RANGE_INTERPOLATED/GEOMETRIC_CENTER, below). This does not
+    touch the anchor path at all — an anchor's own radius is always authoritative
+    when one has been resolved.
     """
     import httpx
     from app.config import settings
@@ -594,11 +617,21 @@ async def _geocode_location(
             lat, lng = float(loc["lat"]), float(loc["lng"])
             dist_km = _geo_dist_km(ref_lat, ref_lon, lat, lng)
             if dist_km > max_km:
-                logger.warning(
+                # server-26#159: the node-distance bound is a proxy for "is
-                    f"Geocoding rejected '{location_str}' → ({lat:.4f}, {lng:.4f}) "
+                # this plausible" that only makes sense when the node's own
-                    f"— {dist_km:.1f}km from {ref_label} exceeds {max_km:.1f}km"
+                # position is our best guess at the area — never when the
+                # query already names a different region on its own terms.
+                if not (ref_label == "node" and trust_named_region):
+                    logger.warning(
+                        f"Geocoding rejected '{location_str}' → ({lat:.4f}, {lng:.4f}) "
+                        f"— {dist_km:.1f}km from {ref_label} exceeds {max_km:.1f}km"
+                    )
+                    return None
+                logger.info(
+                    f"Geocoding '{location_str}' → ({lat:.4f}, {lng:.4f}) is "
+                    f"{dist_km:.1f}km from the receiving node, past {max_km:.1f}km — "
+                    f"accepted anyway: the query names its own region, not the node's"
                 )
-                return None
             coords = {"lat": lat, "lng": lng}
             logger.info(
                 f"Geocoded '{location_str}' → {coords} "
@@ -624,6 +657,43 @@ def _municipality_from_tg(tg_name: Optional[str]) -> Optional[str]:
     return cleaned
 
 
+def _location_query_parts(
+    location: str,
+    tg_area: dict,
+    talkgroup_name: Optional[str],
+    node_state: str,
+    node_county: str,
+) -> tuple[list[str], bool]:
+    """
+    Build the geocode query parts for `location`, plus whether the query names
+    a region the *talkgroup itself* covers (operator-set area_context, or a
+    municipality parsed from the talkgroup's own name) rather than one guessed
+    from wherever the receiving node happens to sit (server-26#159).
+
+    That distinction matters downstream: `_geocode_location`'s node-distance
+    sanity check is only a valid proxy for "is this plausible" when the node's
+    own position is the best guess we have at the area. A citywide/patched
+    feed ("New York City - NYPD Citywide 2 Patch") names its own coverage area
+    right in the talkgroup name — grafting the node's own county onto that
+    (Ossining-style: valid when the feed genuinely is local to the node,
+    actively wrong when it names a distant region of its own) would make the
+    query self-contradictory, so the node's COUNTY is used only when nothing
+    better names the place. The node's STATE is coarse enough to still be
+    correct either way and is kept in both branches.
+    """
+    parts = [location]
+    if area_context.has_place(tg_area):
+        parts += [tg_area[f] for f in area_context.PLACE_FIELDS if tg_area.get(f)]
+        return parts, True
+
+    muni = _municipality_from_tg(talkgroup_name)
+    if muni:
+        parts += [p for p in (muni, node_state) if p]
+    else:
+        parts += [p for p in (node_county, node_state) if p]
+    return parts, muni is not None
+
+
 def _build_transcript_block(transcript: str, segments: Optional[list[dict]]) -> str:
     """Format transcript as numbered transmissions if segments are available."""
     if segments and len(segments) > 1:

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

@@ -522,11 +522,18 @@ async def _run_intelligence_pipeline(
 
         # Correlator also runs for calls with no scenes (unclassified) to attempt
         # talkgroup-based linking even when no transcript could be produced.
-        # Skip when extraction flagged the call — garbage or too-short transcripts
+        # transcript_too_short (<=5 words: "10-8", "show me clear", a unit
-        # carry no signal and would only attach spuriously via the thin path.
+        # check-in) still carries a real transcript and talkgroup — exactly the
+        # brief follow-up/clearance traffic an incident needs, and the thin-path
+        # merge below already requires a same-talkgroup, recently-active
+        # incident before attaching anything, same guard already trusted for
+        # no-transcript calls. Previously excluded here, so these calls never
+        # attached to anything at all. garbage_transcript (Whisper
+        # hallucination) has no real content behind it and stays excluded.
         if not scenes:
             _call_doc = await fstore.doc_get("calls", call_id)
-            if not (_call_doc or {}).get("skip_reason"):
+            skip_reason = (_call_doc or {}).get("skip_reason")
+            if not skip_reason or skip_reason == "transcript_too_short":
                 incident_id = await _correlate_with_consensus(
                     call_id=call_id,
                     node_id=node_id,

drb-c2-core/tests/test_correlator_merge_caps.py

@@ -180,6 +180,55 @@ def test_tactical_thin_call_is_ambiguous_with_two_candidates():
     assert decision["action"] == "orphan"
 
 
+# ---------------------------------------------------------------------------
+# server-26#158: srcaddr identity beats recency guesswork for thin calls
+# ---------------------------------------------------------------------------
+
+def test_thin_call_srcaddr_match_resolves_tier2_ambiguity():
+    """
+    Same fixture as test_tactical_thin_call_is_ambiguous_with_two_candidates —
+    two candidates, tier-2 window, no unit ID parsed (transcript_too_short
+    skipped GPT). Without srcaddr this orphans. With it, the radio that sent
+    the call already touched inc-b, so that's the thread — not a guess.
+    """
+    a = _incident(idle_minutes=3.0, incident_id="inc-a", srcaddrs=["9001"])
+    b = _incident(idle_minutes=4.0, incident_id="inc-b", srcaddrs=["9002"])
+    decision = _run_decision(_ctx(
+        all_active=[a, b], recent=[a, b], talkgroup_name=TACTICAL_TG,
+        call_srcaddr="9002",
+    ))
+    assert decision["action"] == "link"
+    assert decision["matched_incident"]["incident_id"] == "inc-b"
+    assert decision["corr_debug"]["corr_fit_signal"] == "thin_srcaddr_match"
+
+
+def test_thin_call_srcaddr_match_overrides_recency_in_tier1():
+    """
+    Both candidates are inside the 30s conversational window, where recency
+    alone would pick inc-a (more recently updated) even though the radio that
+    sent this call has only ever touched inc-b — the exact busy-channel,
+    two-concurrent-incidents misattach server-26#158 was filed for.
+    """
+    a = _incident(idle_minutes=0.1, incident_id="inc-a", srcaddrs=["9001"])
+    b = _incident(idle_minutes=0.2, incident_id="inc-b", srcaddrs=["9002"])
+    decision = _run_decision(_ctx(
+        all_active=[a, b], recent=[a, b], call_srcaddr="9002",
+    ))
+    assert decision["action"] == "link"
+    assert decision["matched_incident"]["incident_id"] == "inc-b"
+
+
+def test_thin_call_with_no_srcaddr_match_falls_back_to_recency():
+    """A radio ID that matches nothing on this talkgroup behaves exactly as
+    before — no regression for the ordinary case."""
+    a = _incident(idle_minutes=0.1, incident_id="inc-a", srcaddrs=["9001"])
+    decision = _run_decision(_ctx(
+        all_active=[a], recent=[a], call_srcaddr="unrelated-radio",
+    ))
+    assert decision["action"] == "link"
+    assert decision["corr_debug"]["corr_fit_signal"] == "thin_recency"
+
+
 def test_call_with_unit_overlap_does_attach():
     """
     Positive control: real evidence still links. Carrying units also means the

drb-c2-core/tests/test_geocode_named_region.py

@@ -0,0 +1,185 @@
+"""
+server-26#159: a citywide/patched feed can be received far from its own
+coverage area — "New York City - NYPD Citywide 2 Patch" was ~56km from the
+receiving node, well past geocode_max_km (40km). Real, correctly-geocoded
+addresses on that talkgroup were rejected by intelligence._geocode_location's
+node-distance sanity check every time, so location_coords never populated for
+the whole system: location_proximity correlation was permanently dead there,
+and the same real event reported at two nearby addresses two minutes apart
+became two separate incidents instead of one.
+
+`trust_named_region` fixes this narrowly: the node-distance check is a proxy
+for "is this plausible" that only makes sense when the node's own position is
+the best guess we have at the area. It must not apply when the query already
+names a different region on its own terms (operator-set area_context, or a
+municipality parsed straight from the talkgroup's own name) — and it must
+never touch the anchor path, whose own radius is always authoritative.
+"""
+from unittest.mock import patch
+
+import pytest
+
+import app.internal.intelligence as intel
+from app.config import settings
+
+
+def _maps_result(lat: float, lng: float, location_type: str = "ROOFTOP"):
+    payload = {
+        "status": "OK",
+        "results": [{
+            "geometry": {
+                "location": {"lat": lat, "lng": lng},
+                "location_type": location_type,
+            },
+        }],
+    }
+
+    class _Resp:
+        def raise_for_status(self): pass
+        def json(self): return payload
+
+    class _Client:
+        async def __aenter__(self): return self
+        async def __aexit__(self, *a): return False
+        async def get(self, *a, **k): return _Resp()
+
+    return patch("httpx.AsyncClient", lambda *a, **k: _Client())
+
+
+@pytest.fixture(autouse=True)
+def _api_key():
+    # intelligence.py imports settings locally per-function (`from app.config
+    # import settings`), which binds the same cached singleton — patching the
+    # module-level object here reaches it, but `intel.settings` itself does
+    # not exist as an attribute.
+    with patch.object(settings, "google_maps_api_key", "test-key"):
+        yield
+
+
+# Node at (0, 0); result at (1, 0) is ~111km away — well past the 40km default.
+NODE_LAT, NODE_LON = 0.0, 0.0
+FAR_LAT, FAR_LNG = 1.0, 0.0
+
+
+@pytest.mark.asyncio
+async def test_named_region_geocode_accepted_beyond_node_distance():
+    with _maps_result(FAR_LAT, FAR_LNG):
+        coords = await intel._geocode_location(
+            "1108 Jackson Avenue, New York City - NYPD Citywide 2 Patch",
+            node_lat=NODE_LAT, node_lon=NODE_LON,
+            trust_named_region=True,
+        )
+    assert coords == {"lat": FAR_LAT, "lng": FAR_LNG}
+
+
+@pytest.mark.asyncio
+async def test_local_geocode_still_rejected_beyond_node_distance_without_named_region():
+    """Regression guard: a bare street name with no named region still uses
+    the node as its only plausibility check, exactly as before this fix."""
+    with _maps_result(FAR_LAT, FAR_LNG):
+        coords = await intel._geocode_location(
+            "Main Street",
+            node_lat=NODE_LAT, node_lon=NODE_LON,
+            trust_named_region=False,
+        )
+    assert coords is None
+
+
+@pytest.mark.asyncio
+async def test_anchor_path_ignores_trust_named_region():
+    """The anchor's own radius is always authoritative — trust_named_region
+    is only a statement about the node fallback, never a way to widen an
+    anchor that was itself deliberately sized to discriminate."""
+    anchor = {"lat": NODE_LAT, "lng": NODE_LON, "radius_km": 10.0}
+    with _maps_result(FAR_LAT, FAR_LNG):
+        coords = await intel._geocode_location(
+            "1108 Jackson Avenue, New York City - NYPD Citywide 2 Patch",
+            node_lat=NODE_LAT, node_lon=NODE_LON,
+            anchor=anchor, trust_named_region=True,
+        )
+    assert coords is None
+
+
+@pytest.mark.asyncio
+async def test_named_region_geocode_within_node_distance_is_unaffected():
+    """A close result is accepted the same way regardless of the flag."""
+    near_lat, near_lng = 0.05, 0.0  # ~5.5km from the node
+    with _maps_result(near_lat, near_lng):
+        coords = await intel._geocode_location(
+            "Main Street, Ossining, New York",
+            node_lat=NODE_LAT, node_lon=NODE_LON,
+            trust_named_region=True,
+        )
+    assert coords == {"lat": near_lat, "lng": near_lng}
+
+
+@pytest.mark.asyncio
+async def test_imprecise_result_still_rejected_regardless_of_trust():
+    """trust_named_region relaxes the distance check only — the location_type
+    precision filter (server-26#37) still applies unconditionally."""
+    with _maps_result(FAR_LAT, FAR_LNG, location_type="APPROXIMATE"):
+        coords = await intel._geocode_location(
+            "1108 Jackson Avenue, New York City - NYPD Citywide 2 Patch",
+            node_lat=NODE_LAT, node_lon=NODE_LON,
+            trust_named_region=True,
+        )
+    assert coords is None
+
+
+# ---------------------------------------------------------------------------
+# _location_query_parts — pure query assembly, no HTTP involved
+# ---------------------------------------------------------------------------
+
+def test_operator_configured_area_wins_and_is_named_region():
+    parts, named = intel._location_query_parts(
+        "High Street", {"municipality": "Yorktown", "state": "New York"},
+        "Tac 1", node_state="New York", node_county="Westchester",
+    )
+    assert parts == ["High Street", "Yorktown", "New York"]
+    assert named is True
+
+
+def test_local_talkgroup_name_gets_node_state_but_not_node_county():
+    """
+    "Ossining PD" is genuinely local to the node, so appending the node's own
+    state is correct. Its COUNTY is dropped even here — server-26#159's fix
+    applies uniformly once a municipality is derived, since there is no way
+    to tell "local" and "distant-but-node-adjacent" apart from the string
+    alone, and the county was never necessary for a bare municipality name
+    that already disambiguates via the state.
+    """
+    parts, named = intel._location_query_parts(
+        "High Street", {}, "Ossining PD",
+        node_state="New York", node_county="Westchester",
+    )
+    assert parts == ["High Street", "Ossining", "New York"]
+    assert named is True
+
+
+def test_citywide_patched_feed_does_not_get_the_nodes_county_grafted_on():
+    """
+    server-26#159's actual production case: the talkgroup names its own
+    (distant) region, so the node's county (Westchester, ~56km away) must not
+    be appended — it would make the query self-contradictory ("...New York
+    City..., Westchester, New York") and risks degrading the geocode result's
+    precision independently of the distance check this issue also fixes.
+    """
+    parts, named = intel._location_query_parts(
+        "1108 Jackson Avenue", {}, "New York City - NYPD Citywide 2 Patch",
+        node_state="New York", node_county="Westchester",
+    )
+    assert "Westchester" not in parts
+    assert parts == ["1108 Jackson Avenue", "New York City - NYPD Citywide 2 Patch", "New York"]
+    assert named is True
+
+
+def test_uninformative_talkgroup_name_falls_back_to_node_county_and_state():
+    """A tactical channel or bare code gives _municipality_from_tg nothing —
+    the only remaining evidence really is where the node sits, so the
+    original node-county-and-state fallback is preserved for this case."""
+    parts, named = intel._location_query_parts(
+        "High Street", {}, "Tac 1",
+        node_state="New York", node_county="Westchester",
+    )
+    assert parts == ["High Street", "Westchester", "New York"]
+    assert named is False

drb-c2-core/tests/test_upload_short_transcript_correlation.py

@@ -0,0 +1,62 @@
+"""
+server-26#<pending> — a transcript_too_short call (<=5 words: "10-8", "show me
+clear", a unit check-in) never reached correlation at all. upload.py's
+no-scenes fallback (the path that lets a no-transcript call still thin-link
+by talkgroup) explicitly excluded ANY skip_reason, so short-but-real follow-up
+and clearance traffic was permanently unlinkable — not just unextracted by
+GPT, but never even attempted against the fast/thin path that already exists
+for exactly this kind of content-free signal. garbage_transcript (Whisper
+hallucination) has no real content behind it and should stay excluded.
+"""
+from unittest.mock import AsyncMock, patch
+
+import pytest
+
+from app.routers import upload
+
+ALL_ON = {
+    "stt_enabled": True,
+    "correlation_enabled": True,
+    "summaries_enabled": True,
+    "vocabulary_learning_enabled": True,
+    "transcript_correction_enabled": True,
+}
+
+
+async def _run_ingest(skip_reason):
+    with patch("app.internal.feature_flags.get_flags",
+               AsyncMock(return_value=ALL_ON)), \
+         patch("app.internal.firestore.doc_get_cached",
+               AsyncMock(return_value={"system_id": "sys-1", "ai_flags": {}})), \
+         patch.object(upload, "fstore") as fs, \
+         patch.object(upload, "_correlate_with_consensus", AsyncMock(return_value=None)) as corr, \
+         patch("app.internal.transcription.transcribe_call",
+               AsyncMock(return_value=("10-8", []))), \
+         patch("app.internal.intelligence.extract_scenes", AsyncMock(return_value=[])), \
+         patch("app.internal.alerter.check_and_dispatch", AsyncMock()):
+        fs.doc_get = AsyncMock(return_value={"skip_reason": skip_reason} if skip_reason else {})
+        fs.doc_set = AsyncMock()
+        await upload._run_intelligence_pipeline(
+            call_id="call-1", node_id="node-1", system_id="sys-1",
+            talkgroup_id=101, talkgroup_name="PD Dispatch",
+            gcs_uri="gs://bucket/call-1.mp3",
+        )
+    return corr
+
+
+@pytest.mark.asyncio
+async def test_transcript_too_short_now_attempts_correlation():
+    corr = await _run_ingest("transcript_too_short")
+    corr.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_garbage_transcript_still_skips_correlation():
+    corr = await _run_ingest("garbage_transcript")
+    corr.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_no_skip_reason_still_attempts_correlation():
+    corr = await _run_ingest(None)
+    corr.assert_awaited_once()