v0.10.8: discriminate Nats-Msg-Id by event.category to prevent incident+perimeter dedup collision (#96)

JetStream's 2-min per-stream duplicate window keys on Nats-Msg-Id alone --
the subject is NOT part of the dedup key. So when wfigs_incidents and
wfigs_perimeters published envelopes for the same Summit Creek IRWIN
{6B1C6EB1-30F7-4613-9C58-4801DC8FD822} within seconds of each other on
2026-06-07, the second publish was silently dropped: supervisor logged
'Published event' for both, but CENTRAL_FIRE's last_seq advanced by 25
instead of 26.

Root cause: cloudevents_wire.wrap_event returned event.id verbatim as
msg_id, and both wfigs adapters use the bare IRWIN GUID as event.id.

Fix: synthesize msg_id as f'{event.id}:{event.category}' in wrap_event.
Categories already differ ('fire.incident.X' vs 'fire.perimeter.X') so
this is a natural and load-bearing discriminator. envelope['id'] stays
the bare event.id per CloudEvents spec, so subscribers that key off the
payload id field are unaffected. Adapter-side event.id construction
unchanged.

Backward compatibility: this is a one-time msg_id-shape change. The
JetStream 2-min dedup window is per-(event-id, category) starting now
instead of per-event-id. For events still in the dedup window at the
moment of the deploy under the OLD shape, their first post-deploy
publish under the NEW shape is a fresh msg_id and lands normally -- no
collision risk. Quake and other single-class adapters get a no-op shape
change (their category is constant per adapter, so dedup still catches
genuine same-id duplicates).

Tests:
- test_required_fields_present updated for the new assertion.
- New test_msgid_disambiguates_incident_vs_perimeter_same_guid regression
  guard with the real Summit Creek IRWIN -- assert the two msg_ids differ
  and match the expected shape exactly.
- New parametrize test_msgid_shape_is_id_colon_category over 5
  categories covering multi-class (fire.incident, fire.perimeter) and
  single-class (quake, wx.alert, hydro) adapters.

Full sweep: 1028 passed, 0 failures (+6 from this PR), ruff clean.

Deploy: NOT an output-shape change to envelope payload, no published_ids
flush needed. Squash-merge -> tag v0.10.8 -> pull on central -> restart
central-supervisor. Verify by forcing a Summit Creek republish and
confirming both .incident.id.cassia AND .perimeter.id.cassia land
post-deploy with the same IRWIN.

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

tests/test_models.py

@@ -64,11 +64,17 @@ class TestCloudEventsWire:
     def test_required_fields_present(
         self, sample_event: Event, sample_config: Config
     ) -> None:
-        """Required CloudEvents fields are present."""
+        """Required CloudEvents fields are present.
+
+        v0.10.8: msg_id is now category-discriminated. The CloudEvents
+        envelope ``id`` stays the bare ``event.id`` per CE spec; only the
+        ``Nats-Msg-Id`` header (= the second return value of wrap_event)
+        carries the ``"{event.id}:{event.category}"`` shape.
+        """
         envelope, msg_id = wrap_event(sample_event, sample_config)
 
-        assert msg_id == sample_event.id
-        assert envelope["id"] == sample_event.id
+        assert msg_id == f"{sample_event.id}:{sample_event.category}"
+        assert envelope["id"] == sample_event.id  # CE spec: bare event.id
         assert envelope["source"] == sample_config.cloudevents.source
         assert envelope["type"] == "central.wx.alert.severe_thunderstorm_warning.v1"
         assert envelope["specversion"] == "1.0"
@@ -76,6 +82,77 @@ class TestCloudEventsWire:
         assert envelope["datacontenttype"] == "application/json"
         assert "data" in envelope
 
+    def test_msgid_disambiguates_incident_vs_perimeter_same_guid(
+        self, sample_geo: Geo, sample_config: Config
+    ) -> None:
+        """v0.10.8 regression guard: Summit Creek dedup collision.
+
+        On 2026-06-07 a wfigs poll yielded both an incident envelope and
+        a perimeter envelope for the same IRWIN ``{6B1C6EB1-...}`` within
+        JetStream's 2-min dedup window on CENTRAL_FIRE. Both publishes
+        carried the bare IRWIN as ``Nats-Msg-Id`` (JetStream's dedup key
+        is per-stream, NOT per-subject), so the second publish was
+        silently dropped -- supervisor logs showed 'Published event' for
+        both but only one landed in the stream. The fix adds an
+        ``:{event.category}`` suffix to the msg_id so the two are
+        distinct without changing ``envelope["id"]`` (CE spec) or the
+        adapter-side event.id contract.
+        """
+        irwin = "{6B1C6EB1-30F7-4613-9C58-4801DC8FD822}"  # Summit Creek
+        incident = Event(
+            id=irwin,
+            adapter="wfigs_incidents",
+            category="fire.incident.wildfire",
+            time=datetime(2026, 6, 7, 6, 25, 5, tzinfo=timezone.utc),
+            severity=2,
+            geo=sample_geo,
+            data={},
+        )
+        perimeter = Event(
+            id=irwin,
+            adapter="wfigs_perimeters",
+            category="fire.perimeter.wildfire",
+            time=datetime(2026, 6, 7, 6, 25, 5, tzinfo=timezone.utc),
+            severity=2,
+            geo=sample_geo,
+            data={},
+        )
+
+        _, incident_msgid = wrap_event(incident, sample_config)
+        _, perimeter_msgid = wrap_event(perimeter, sample_config)
+
+        assert incident_msgid != perimeter_msgid
+        assert incident_msgid == f"{irwin}:fire.incident.wildfire"
+        assert perimeter_msgid == f"{irwin}:fire.perimeter.wildfire"
+
+    @pytest.mark.parametrize("category", [
+        "fire.incident.wildfire",
+        "fire.perimeter.wildfire",
+        "wx.alert.severe_thunderstorm_warning",
+        "quake.usgs_quake",
+        "hydro.00060.usgs.13162225",
+    ])
+    def test_msgid_shape_is_id_colon_category(
+        self, sample_geo: Geo, sample_config: Config, category: str
+    ) -> None:
+        """Every wrap_event call returns msg_id == ``f'{id}:{category}'``.
+
+        Covers both multi-class adapters (where the discriminator is
+        load-bearing) and single-class adapters (where it's a no-op for
+        dedup but still consistent shape across the codebase).
+        """
+        event = Event(
+            id=f"test:{category}:42",
+            adapter="test",
+            category=category,
+            time=datetime(2026, 6, 7, 6, 25, 5, tzinfo=timezone.utc),
+            severity=1,
+            geo=sample_geo,
+            data={},
+        )
+        _, msg_id = wrap_event(event, sample_config)
+        assert msg_id == f"{event.id}:{category}"
+
     def test_extension_attributes_lowercase(
         self, sample_event: Event, sample_config: Config
     ) -> None: