Move wfigs wire rendering into notifications/formatters/fire.py and the
full fire state machine into notifications/gating/fire.py, behind the
registry, NO cutover. handle_wfigs builds a canonical dict, calls
decide(), keeps the inline fires INSERT/UPDATE of current_* and the
tombstoned_at stamp unconditional, then branches on is_cutover(...) —
legacy _attach_commit_handles/all-clear path stays byte-identical while
the new path bakes in shadow.
Reproduces every legacy stamp through GateResult.data_patch:
- forward-only acres/containment growth + 8h cooldown gating
- tombstone wildfire_closed all-clear (row exists AND last_broadcast_at
IS NOT NULL) with _severity_override="priority", _dedup_suffix="closed"
- growth _dedup_suffix=f"{acres}|{contained_pct}", _cooldown_suffix=irwin_id
- idempotent commit UPSERT of fires(last_broadcast_*) + event_log flip
- full _location_anchor fallback chain (geocoder_city -> resolve_anchor
-> landclass -> county -> state) preserved in the formatter
Registered under the three explicit categories (wildfire_declared,
wildfire_incident, wildfire_closed) rather than the `fire` toggle, so the
family-fallback does NOT capture the still-deferred FIRMS categories
(wildfire_hotspot/new_ignition/wildfire_growth); a registration test
asserts those resolve elsewhere.
Native env/fires.py deferred (missing IRWIN/cause/landclass, no tombstone
concept); non-cutover so store._emit_event won't run it.
tier-a: 19 new golden+gate-sequence tests; wfigs handler 23/23 preserved;
suite at 34-failure baseline (1571 passed).
Co-authored-by: Matt Johnson <mj@k7zvx.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
shadow_gate called decider(data, now=now), omitting the required
keyword-only `source` arg from the decide(data, *, source, now)
contract. Every decider raised TypeError (swallowed at DEBUG), so the
Phase-1/2 gate-shadow was a silent no-op the entire bake and wrote zero
mismatch records.
- shadow.py:139 -> decider(data, source=source, now=now)
- add TestShadowGateForwardsSource regression guard (fake decider that
asserts source is forwarded; fails pre-fix, passes post-fix)
- retarget TestShadowInertWhenNoDecider at a genuinely-unregistered
category (__no_such_category__); it was using earthquake_event, whose
decider only "passed" before because the signature bug crashed it
Suite at 34-failure baseline.
Co-authored-by: Matt Johnson <mj@k7zvx.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(fixtures): rewrite capture_fixtures.py to use ephemeral push subscribe
Previous script hardcoded stream="CENTRAL" which does not exist — Central
partitions into domain streams (CENTRAL_QUAKE, CENTRAL_SPACE, etc.). It
also called pull_subscribe_bind() without await, making the fetch a no-op.
Fix: mirror the proven CentralConsumer.start() pattern — use
js.subscribe(subject, cb=..., AckPolicy.NONE, no durable) which auto-
discovers the correct stream via the subject filter, identical to how the
live consumer binds. Messages are funnelled through asyncio.Queue with
an idle-timeout to detect drain completion.
Adds live captured fixtures:
- tests/fixtures/quake/ — 3 envelopes (CENTRAL_QUAKE stream, mode=all)
- tests/fixtures/swpc/ — 40 envelopes (CENTRAL_SPACE, mode=all, proton_flux history)
- tests/fixtures/swpc_last/ — 23 envelopes (mode=last: 21 alert variants + kindex + proton_flux)
Avalanche: confirmed empty off-season (CENTRAL_AVY stream, 0 messages).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* refactor(phase1): migrate quake, swpc(Kp+flare), avalanche + staged-cutover gate
First hazards on the source-agnostic formatter+decider path, behind a
staged-cutover gate so deploy = shadow-only (old path still broadcasts; new
path dry-run-diffed) until MESHAI_CUTOVER_CATEGORIES flips a category live.
- formatters/{quake,swpc,avalanche}.py + gating/{quake,swpc,avalanche}.py:
source-agnostic format(event,*,now,budget) + decide(data,*,source,now)->GateResult.
quake (earthquake_event, tier-b: render PAGER + live update-prefix), swpc
(geomagnetic_storm + rf_propagation_alert, Kp+flare only; proton/solar_radiation
_storm stays legacy; geomag 600s window re-homed off the module-global into
gating/swpc with a deferred stamp; tier-b scale-based severity), avalanche
(avalanche_warning/watch; centralseverity->NAADS 1-5 remap; synthetic fixtures
off-season).
- central/{quake,swpc,avy}_handler.py bridges: cutover -> new decide()+canonical
data; not-cutover -> exact legacy behavior. env/{usgs_quake,swpc,avalanche}.py
emit canonical Event.data (avalanche stops precomposing). env/store.py generic
native decider hook (cutover-gated).
- notifications/cutover.py (is_cutover via MESHAI_CUTOVER_CATEGORIES); composer
dispatch + shadow hooks are cutover-aware (shadow no-ops once a category is live).
- scripts/capture_fixtures.py fixed (per-domain streams e.g. CENTRAL_QUAKE, await
bind); real quake/swpc fixtures captured; avalanche synthetic.
Tests: +~150 (quake/swpc/avalanche parity+cross-source+gate-sequence+cutover);
0 new failures (34 baseline, 1426 passed).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Matt Johnson <mj@k7zvx.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Foundation for making all hazard formatting+gating source-agnostic. ZERO
behavior change — the formatter/decider registries are empty (get_formatter/
get_decider return None → existing precomposed/Mode-B path preserved), and the
shadow comparator is off unless MESHAI_SHADOW_CATEGORIES is set.
- notifications/formatters/ (registry+dispatch with family fallback), gating/
(GateResult + deferred-commit contract), both empty registries.
- notifications/clock.py determinism seam; route wfigs/quake/nws gating time
reads through it (identical values) so goldens can freeze time.
- formatters/_budget.py = copy of central/budget.py; central/budget.py is now a
re-export shim (import-smoke test guards it).
- compose_mesh_message consults the registry first (verbatim, no Mode-B re-cap),
falls back to legacy; _resolve_budget injects per-category budget.
- notifications/shadow.py + two DRY-RUN hooks (consumer._normalize, dispatcher
render): compute the new result and diff-log SHADOW_MISMATCH JSONL, but NEVER
commit/emit/write tables and always broadcast the OLD result. Inert by default.
- tests/harness (pinned_time/pinned_tz, byte-golden, gate-sequence) +
scripts/capture_fixtures.py (ephemeral read-only NATS capture); tzdata pinned.
Tests: +60 (18 scaffold + 42 harness/shadow); 0 new failures (34 baseline).
Co-authored-by: Matt Johnson <mj@k7zvx.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
