meshai/work/tests/test_firms_native_fusion.py

389 lines
17 KiB
Python
Raw Permalink Normal View History

"""Native FIRMS fire-fusion tests — the source-agnostic ingest entrypoint.
Covers the wiring that lets the native ``env/firms.py`` adapter feed
locally-fetched NASA FIRMS pixels into the SAME attribution/fusion pipeline the
Central handler uses, via ``central.firms_handler.ingest_hotspot_pixel``:
1. ``ingest_hotspot_pixel`` drives growth / spotting / halt from canonical
pixels (and, F3, curated ``unattributed_hotspot_cluster`` "possible new
fire" broadcasts) and returns ``(wire, data)`` — and NEVER a raw per-pixel
hotspot / new_ignition broadcast.
2. ``env/firms.py`` tick() (CSV fetch monkeypatched) feeds those pixels through
the shared engine, emits the fusion Events, and still returns None for raw
hotspots.
3. A guard that the Central ``handle_firms`` path is unchanged by the
extraction (storage side-effects + growth wire/stamps identical).
Mirrors the driving style of test_firms_refactor.py / test_fire_tracker_phase*.
"""
from __future__ import annotations
import math
import time
import uuid
from datetime import datetime, timezone
import pytest
_MI_PER_DEG_LAT = 69.0
# F3: unattributed_hotspot_cluster is a CURATED fusion output (a "possible new
# fire"), not a raw per-pixel broadcast. Raw hotspots are still never emitted.
_FUSION_CATS = {"wildfire_growth", "wildfire_spotting", "wildfire_halted",
"unattributed_hotspot_cluster"}
_RAW_CATS = {"wildfire_hotspot", "new_ignition"}
# ── isolation (real-DB, mirrors test_firms_refactor) ─────────────────────────
@pytest.fixture(autouse=True)
def _isolate_db(tmp_path, monkeypatch):
db_path = str(tmp_path / f"meshai-{uuid.uuid4().hex}.sqlite")
monkeypatch.setenv("MESHAI_DB_PATH", db_path)
from meshai.persistence import db as pdb
pdb.close_thread_connection()
pdb._initialised.discard(db_path)
from meshai.persistence import init_db
init_db(db_path)
try:
from meshai.adapter_config import adapter_config as _ac
_ac.invalidate()
except Exception:
pass
yield db_path
pdb.close_thread_connection()
pdb._initialised.discard(db_path)
@pytest.fixture(autouse=True)
def _no_cutover(monkeypatch):
"""Default deploy state: nothing cut over -> legacy stamps, direct emit."""
monkeypatch.delenv("MESHAI_CUTOVER_CATEGORIES", raising=False)
from meshai.notifications.cutover import _clear_cache
_clear_cache()
yield
_clear_cache()
# ── helpers ──────────────────────────────────────────────────────────────────
def _seed_fire(*, irwin_id, lat, lon, name="Stub Fire", **cols):
from meshai.persistence import get_db
conn = get_db()
base = {"irwin_id": irwin_id, "incident_name": name, "lat": lat, "lon": lon,
"last_event_at": int(time.time())}
base.update(cols)
keys = ",".join(base)
ph = ",".join("?" * len(base))
conn.execute(f"INSERT INTO fires({keys}) VALUES ({ph})", tuple(base.values()))
def _acq_epoch(acq_date: str, acq_time: str) -> int:
return int(datetime.strptime(f"{acq_date} {acq_time.zfill(4)}",
"%Y-%m-%d %H%M")
.replace(tzinfo=timezone.utc).timestamp())
def _pixel(*, lat, lon, acq_date="2026-06-06", acq_time="1200",
frp=20.0, confidence="high", brightness=320.0, satellite="N20"):
"""Build a canonical FIRMS pixel dict."""
return {
"lat": lat, "lon": lon, "frp": frp, "confidence": confidence,
"brightness": brightness, "satellite": satellite,
"acq_epoch": _acq_epoch(acq_date, acq_time),
}
def _offset_mi(lat, lon, north_mi, east_mi):
dlat = north_mi / _MI_PER_DEG_LAT
dlon = east_mi / (_MI_PER_DEG_LAT * math.cos(math.radians(lat)))
return lat + dlat, lon + dlon
def _feed(pixel, *, now):
chore(central-ripout 2d-i): relocate the fire engine + renderer out of central/ (#164) Moves the last two live files out of the retired-Central folder. central/ is now EMPTY and deleted entirely (incl. __init__.py). - fire-fusion engine (ingest_hotspot_pixel + the growth/cluster/spotting engine) → env/fire_fusion.py, next to its sole consumer env/firms.py. - wildfire text renderer (_render + its live helpers) → env/fire_render.py. THE COUPLING RESOLVED: firms_handler._handle_pass_boundary had a LAZY import inside a function body — `from meshai.central.wfigs_handler import _render` — sitting on the live FIRMS fire-growth path. It is now a normal top-of-file import (`from meshai.env.fire_render import _render`), visible and greppable. PURE MOVE — no behavior change: - The parity oracle (test_fire_refactor.py) PASSES UNCHANGED (only its import paths updated) — proving the fire wire output is byte-for-byte identical before and after. Fire alerts say exactly what they said. - handle_firms / handle_wfigs (dead entrypoints, only caller was the deleted consumer.py) were KEPT and moved rather than dropped — the wording-cleanup PR removes them deliberately. "When unsure, keep." - fire_render.py's geo-helper imports still point at central_normalizer — that file's split is a SEPARATE PR; carried the imports along, did not touch it. Consumers + test import paths rewired. Full suite: 2059 passed, 0 failed. Co-authored-by: Matt Johnson <mj@k7zvx.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-17 16:31:48 -06:00
from meshai.env.fire_fusion import ingest_hotspot_pixel
return ingest_hotspot_pixel(pixel, now=now)
def _assert_no_raw(broadcasts):
"""Every returned broadcast must be a fusion category; never a raw one."""
for wire, data in broadcasts:
cat = data.get("category")
assert cat in _FUSION_CATS, f"unexpected non-fusion broadcast: {cat!r}"
assert cat not in _RAW_CATS
# ═════════════════════════════════════════════════════════════════════════════
# 1. ingest_hotspot_pixel — growth / spotting / halt
# ═════════════════════════════════════════════════════════════════════════════
class TestIngestGrowth:
def test_two_pass_growth_broadcasts(self):
center_lat, center_lon = 42.0, -114.0
_seed_fire(irwin_id="ID-NG", lat=center_lat, lon=center_lon,
name="Pine Gulch")
# Pass A: 5 pixels around the anchor (acq 12:00-12:04 = one pass bucket).
for i in range(5):
out = _feed(_pixel(lat=center_lat + 0.0001 * i,
lon=center_lon + 0.0001 * (i - 2),
acq_time=f"12{i:02d}", frp=20.0 + i),
now=1780747200 + i)
_assert_no_raw(out)
assert out == [], "pass-A pixels must not broadcast (no boundary yet)"
# Pass B: one pixel 1 mi north (18:00 = different pass bucket) -> growth.
pass_b_lat = center_lat + (1.0 / _MI_PER_DEG_LAT)
out = _feed(_pixel(lat=pass_b_lat, lon=center_lon, acq_time="1800",
frp=22.0), now=1780768800)
_assert_no_raw(out)
assert len(out) == 1
wire, data = out[0]
assert data["category"] == "wildfire_growth"
assert data["_severity_override"] == "immediate"
assert wire.startswith("🔥 Pine Gulch")
assert "Moving N" in wire
class TestIngestSpotting:
def _seed_hex_pass_a(self, irwin_id, center_lat, center_lon,
start_now=1780747200):
_seed_fire(irwin_id=irwin_id, lat=center_lat, lon=center_lon,
name=irwin_id)
for i in range(6):
angle = i * math.pi / 3
la = center_lat + (0.5 / _MI_PER_DEG_LAT) * math.sin(angle)
cos_lat = math.cos(math.radians(center_lat))
lo = center_lon + (0.5 / (_MI_PER_DEG_LAT * cos_lat)) * math.cos(angle)
out = _feed(_pixel(lat=la, lon=lo, acq_time=f"12{i * 2:02d}"),
now=start_now + i)
_assert_no_raw(out)
def test_spotting_broadcasts(self):
center_lat, center_lon = 43.0, -115.0
self._seed_hex_pass_a("ID-NS", center_lat, center_lon)
# Pass B pixel ~2 mi NE of the perimeter -> spotting.
sp_lat, sp_lon = _offset_mi(center_lat, center_lon,
north_mi=2.0 / math.sqrt(2),
east_mi=2.0 / math.sqrt(2))
out = _feed(_pixel(lat=sp_lat, lon=sp_lon, acq_time="1800"),
now=1780768800)
_assert_no_raw(out)
assert len(out) == 1
wire, data = out[0]
assert data["category"] == "wildfire_spotting"
fix(firms): repair the FIRMS fire-fusion Event contract (issues #117-#119) (#120) Three independent bugs kept firms_handler's growth/spotting/halt/cluster fusion decisions from reaching a correct mesh Event: - #117: consumer._normalize() computed `category` from the raw Central category BEFORE the per-adapter handler ran and never re-read data["category"] afterward, so every firms_handler category stamp was a silent no-op. Now re-read post-dispatch, validated against the known category registry (unrecognized overrides are logged and ignored). - #118: consumer.py only ever honors data["_severity_override"], but firms_handler's halt/spotting/cluster sites stamped the plain data["severity"] key instead (only growth used the right key). Switched all three sites to `_severity_override` for one consistent contract. This is severity plumbing only -- it does not change which events fire. - #119: FirePacer's gate only matched source in ("fires","wfigs") at severity=="priority", so FIRMS fusion broadcasts (source="firms", growth/spotting at "immediate") never reached the pacer. Broadened the gate to cover "firms" + {"priority","immediate"}, and gave FirePacer head-of-line insertion so an "immediate" event is never stuck behind already-queued "priority" events. Still unbounded/never-drops. Cluster detection is left exactly as main ships it: live, always on, no toggle (PR #73's curated new-fire cluster broadcasts with cold-start silent-seeding). Only its severity-override key changes, under #118. Updated existing tests that asserted the old (buggy) data["severity"] contract, and added tests/test_firms_fusion_event_contract.py covering all three fixes end-to-end through consumer._normalize()/_handle() and FirePacer directly. Co-authored-by: Matt Johnson <mj@k7zvx.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-11 17:41:40 -06:00
assert data["_severity_override"] == "immediate"
assert wire.startswith("🔥 Possible spotting ")
# Eager latch stamped during ingest (not-cutover legacy path).
from meshai.persistence import get_db
latch = get_db().execute(
"SELECT last_spotting_broadcast_at FROM fires WHERE irwin_id=?",
("ID-NS",)).fetchone()[0]
assert latch == 1780768800.0
class TestIngestHalt:
def test_halt_broadcasts_for_idle_fire(self):
now = 1780768800
idle_at = now - 14 * 3600
from meshai.persistence import get_db
get_db().execute(
"INSERT INTO fires(irwin_id, incident_name, lat, lon, "
"last_event_at, last_pass_id, last_pass_at) VALUES (?,?,?,?,?,?,?)",
("ID-NH", "Cold Fire", 42.5, -114.5, int(idle_at),
"N20-329627", float(idle_at)),
)
# A fresh pixel FAR from the idle fire -> no attribution, cluster dead,
# halt detector fires for the idle fire.
out = _feed(_pixel(lat=45.0, lon=-118.0, acq_time="1800"), now=now)
_assert_no_raw(out)
assert len(out) == 1
wire, data = out[0]
assert data["category"] == "wildfire_halted"
fix(firms): repair the FIRMS fire-fusion Event contract (issues #117-#119) (#120) Three independent bugs kept firms_handler's growth/spotting/halt/cluster fusion decisions from reaching a correct mesh Event: - #117: consumer._normalize() computed `category` from the raw Central category BEFORE the per-adapter handler ran and never re-read data["category"] afterward, so every firms_handler category stamp was a silent no-op. Now re-read post-dispatch, validated against the known category registry (unrecognized overrides are logged and ignored). - #118: consumer.py only ever honors data["_severity_override"], but firms_handler's halt/spotting/cluster sites stamped the plain data["severity"] key instead (only growth used the right key). Switched all three sites to `_severity_override` for one consistent contract. This is severity plumbing only -- it does not change which events fire. - #119: FirePacer's gate only matched source in ("fires","wfigs") at severity=="priority", so FIRMS fusion broadcasts (source="firms", growth/spotting at "immediate") never reached the pacer. Broadened the gate to cover "firms" + {"priority","immediate"}, and gave FirePacer head-of-line insertion so an "immediate" event is never stuck behind already-queued "priority" events. Still unbounded/never-drops. Cluster detection is left exactly as main ships it: live, always on, no toggle (PR #73's curated new-fire cluster broadcasts with cold-start silent-seeding). Only its severity-override key changes, under #118. Updated existing tests that asserted the old (buggy) data["severity"] contract, and added tests/test_firms_fusion_event_contract.py covering all three fixes end-to-end through consumer._normalize()/_handle() and FirePacer directly. Co-authored-by: Matt Johnson <mj@k7zvx.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-11 17:41:40 -06:00
assert data["_severity_override"] == "routine"
assert wire == "🔥 Cold Fire no growth in 14h"
latch = get_db().execute(
"SELECT halt_broadcast_at FROM fires WHERE irwin_id=?",
("ID-NH",)).fetchone()[0]
assert latch == float(now)
class TestIngestNeverRaw:
def test_lone_pixel_no_fire_yields_nothing(self):
# Stored, attributed to nothing, below the cluster threshold, no idle
# fire -> returns []. A raw hotspot is NEVER emitted on any path.
out = _feed(_pixel(lat=44.4, lon=-116.2, acq_time="1300"),
now=1780750000)
assert out == []
def test_dense_unattributed_cluster_broadcasts_curated(self):
# F3: several unattributed pixels close together form a CURATED cluster
# ("possible new fire"). Non-seed path (no cold-start), so it broadcasts:
# a wire fires on the 3rd pixel (min_pixels=3), the first 3 are stamped
# so they can't re-fire, and the next 3 form a fresh cluster -> a 2nd
# wire on the 6th pixel. NEVER a raw per-pixel broadcast.
base_lat, base_lon = 44.0, -116.0
produced = []
for i in range(6):
out = _feed(_pixel(lat=base_lat + 0.001 * i, lon=base_lon,
acq_time=f"13{i:02d}"), now=1780750000 + i)
_assert_no_raw(out)
produced.extend(out)
assert len(produced) == 2, f"expected 2 curated cluster wires: {produced}"
for wire, data in produced:
assert wire.startswith("🔥 Possible new fire:")
assert data["category"] == "unattributed_hotspot_cluster"
fix(firms): repair the FIRMS fire-fusion Event contract (issues #117-#119) (#120) Three independent bugs kept firms_handler's growth/spotting/halt/cluster fusion decisions from reaching a correct mesh Event: - #117: consumer._normalize() computed `category` from the raw Central category BEFORE the per-adapter handler ran and never re-read data["category"] afterward, so every firms_handler category stamp was a silent no-op. Now re-read post-dispatch, validated against the known category registry (unrecognized overrides are logged and ignored). - #118: consumer.py only ever honors data["_severity_override"], but firms_handler's halt/spotting/cluster sites stamped the plain data["severity"] key instead (only growth used the right key). Switched all three sites to `_severity_override` for one consistent contract. This is severity plumbing only -- it does not change which events fire. - #119: FirePacer's gate only matched source in ("fires","wfigs") at severity=="priority", so FIRMS fusion broadcasts (source="firms", growth/spotting at "immediate") never reached the pacer. Broadened the gate to cover "firms" + {"priority","immediate"}, and gave FirePacer head-of-line insertion so an "immediate" event is never stuck behind already-queued "priority" events. Still unbounded/never-drops. Cluster detection is left exactly as main ships it: live, always on, no toggle (PR #73's curated new-fire cluster broadcasts with cold-start silent-seeding). Only its severity-override key changes, under #118. Updated existing tests that asserted the old (buggy) data["severity"] contract, and added tests/test_firms_fusion_event_contract.py covering all three fixes end-to-end through consumer._normalize()/_handle() and FirePacer directly. Co-authored-by: Matt Johnson <mj@k7zvx.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-11 17:41:40 -06:00
assert data["_severity_override"] == "priority"
# ═════════════════════════════════════════════════════════════════════════════
# 2. env/firms.py tick() — shared engine + raw-hotspot suppression
# ═════════════════════════════════════════════════════════════════════════════
def _adapter(source="VIIRS_NOAA20_NRT"):
from unittest.mock import MagicMock
from meshai.env.firms import FIRMSAdapter
cfg = MagicMock()
cfg.map_key = "test-key"
cfg.source = source
cfg.bbox = [-117, 42, -114, 44]
cfg.day_range = 1
cfg.tick_seconds = 0
cfg.confidence_min = "nominal"
cfg.proximity_km = 10.0
a = FIRMSAdapter(cfg, region_anchors=[], fires_adapter=None)
a._last_tick = 0.0
return a
def _csv(rows):
header = "latitude,longitude,bright_ti4,confidence,frp,acq_date,acq_time"
lines = [header]
for r in rows:
lines.append("{lat},{lon},{b},{c},{frp},{d},{t}".format(
lat=r["lat"], lon=r["lon"], b=r.get("b", 320.0),
c=r.get("c", "high"), frp=r.get("frp", 20.0),
d=r.get("d", "2026-06-06"), t=r.get("t", "1200")))
return "\n".join(lines)
class _FakeResp:
def __init__(self, text):
self._text = text
def read(self):
return self._text.encode("utf-8")
def __enter__(self):
return self
def __exit__(self, *a):
return False
def _patch_fetch(monkeypatch, csv_text):
monkeypatch.setattr("meshai.env.firms.urlopen",
lambda *a, **k: _FakeResp(csv_text))
class TestAdapterTickFusion:
def _growth_rows(self, center_lat, center_lon):
# Anchor acq times to NOW (native path uses now=time.time()): pass A
# ~6h ago, pass B ~now. Both are well inside the 12h halt gate, so the
# growing fire never looks idle -- only the growth broadcast fires.
# (6h apart => distinct pass buckets => a real pass boundary.)
now_dt = datetime.now(timezone.utc)
pass_a = now_dt - __import__("datetime").timedelta(hours=6)
rows = []
for i in range(5):
t = pass_a + __import__("datetime").timedelta(minutes=i)
rows.append({"lat": center_lat + 0.0001 * i,
"lon": center_lon + 0.0001 * (i - 2),
"d": t.strftime("%Y-%m-%d"), "t": t.strftime("%H%M"),
"frp": 20.0 + i})
rows.append({"lat": center_lat + 1.0 / _MI_PER_DEG_LAT,
"lon": center_lon,
"d": now_dt.strftime("%Y-%m-%d"),
"t": now_dt.strftime("%H%M"), "frp": 22.0})
return rows
def test_tick_emits_growth_and_suppresses_raw(self, monkeypatch):
center_lat, center_lon = 42.0, -114.0
_seed_fire(irwin_id="ID-TG", lat=center_lat, lon=center_lon,
name="Pine Gulch")
_patch_fetch(monkeypatch, _csv(self._growth_rows(center_lat, center_lon)))
a = _adapter()
# Steady-state growth routing (not cold start): mark the adapter past its
# first-fetch silent-seed so this tick's growth broadcasts. Cold-start
# fusion suppression is covered in test_firms_cluster_f3.
a._firms_seeded = True
assert a.tick() is True
evts = a.get_events()
raw = [e for e in evts if e.get("source") == "firms"]
fusion = [e for e in evts if e.get("source") == "firms_fusion"]
# Raw pixels present (LLM context) but every one renders None.
assert raw, "raw hotspots should still be cached for LLM context"
for r in raw:
assert a.to_event(r) is None
# Exactly one fusion broadcast, and it renders a real growth Event.
assert len(fusion) == 1
ev = a.to_event(fusion[0])
assert ev is not None
assert ev.category == "wildfire_growth"
assert ev.summary.startswith("🔥 Pine Gulch")
assert ev.data.get("_meshai_precomposed") is True
def test_tick_emits_through_store_emit_event(self, monkeypatch):
"""End-to-end: the fusion Event reaches the pipeline bus via
store._emit_event; raw hotspots never do."""
from unittest.mock import MagicMock
from meshai.env.store import EnvironmentalStore
center_lat, center_lon = 42.0, -114.0
_seed_fire(irwin_id="ID-TS", lat=center_lat, lon=center_lon,
name="Pine Gulch")
_patch_fetch(monkeypatch, _csv(self._growth_rows(center_lat, center_lon)))
a = _adapter()
bus = MagicMock()
store = EnvironmentalStore.__new__(EnvironmentalStore)
store._events = {}
store._event_bus = bus
feat(env): native adapters broadcast only newly-RECEIVED items (no backlog) (#43) Replace the native path's "scan accumulated state + suppress what we've already broadcast" model with "broadcast only what newly arrived from the API this poll." Storage is unchanged (self._events + firms_pixels etc. are populated for EVERY received item, so the LLM/get_active backlog is intact); only the BROADCAST decision changes. - env/store.py: per-adapter in-memory seen-set (_seen) + _seeded. First data-bearing poll for an adapter seeds keys and emits NOTHING (that batch is pre-existing backlog); later polls emit only keys not seen before. Restart => empty sets => next poll re-seeds silently. Structurally impossible to broadcast backlog on cold start / restart / re-enable. Key = external_id -> event_id -> content hash, namespaced per adapter. self._events[key]=evt still runs unconditionally (storage preserved). - Fixes the ~175 (roads511) / ~782 (wzdx) cold-start bursts AND the latent quake/nws version (they only looked safe because Central pre-populated their broadcast tables). - env/satpass.py: broadcast on AOS IMMINENCE (now < aos <= now+lead, broadcast_lead_seconds default 3600), future-only; window_hours still governs prediction depth. Strict norad_ids post-filter + fixed _parse_norad_ids char-iteration bug (cause of GOES/METEOR leak). - Central path + broadcast-state tables untouched (native-only gate). 13 new tests; full suite at 10-failure baseline (1697 passed). Co-authored-by: Matt Johnson <mj@k7zvx.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-05 14:38:40 -06:00
# Received-delta gate state. This test exercises STEADY-STATE fusion
# routing (fusion reaches the bus, raw hotspots never do), so mark firms
# as already past its first-poll seed — otherwise the store's
# received-delta gate would (correctly) seed this first batch silently.
# First-poll seed-silence is covered by test_store_received_delta.
fix(env): leak-proof received-delta — durable persistent baseline + non-empty seed guard (#44) Two live backlog-broadcast leaks traced to the in-memory first-poll seed: (1) incremental-fetch adapters (wzdx: registry tick [0 events] then feeds tick [many]) got marked _seeded on the EMPTY first tick, so the real batch next tick all looked "new" and broadcast; (2) in-memory seed lost on restart. Fix — durable baseline + guard: - _seed_from_persistent() at store init: pre-load already-received item keys from the persistent hazard tables into self._seen, so nothing ever received can re-broadcast (immune to fetch staging + restart). Only sources whose native emit key PROVABLY equals a persistent key are durably seeded: wzdx (traffic_events.external_id) + usgs_quake (quake_events.event_id). Resilient (per-table try/except; missing table -> skip). - _seen_key() now namespaces by evt["source"] (matches persistent tables), via shared _key_ext/_key_eid helpers used by both seed and live emit so they can't drift. - non-empty-seed guard: _ingest marks only sources that carried >=1 event this poll as _seeded -> an empty first tick can never seed-then-leak. This is the root-cause fix; covers all adapters (roads511/traffic fetch atomically per tick, so the guard fully protects them). - storage untouched (self._events populated for every event); Central path/deciders untouched. Live-DB verified: seed pre-loads 784 wzdx + 8 quake keys -> a live wzdx poll of 784 known zones broadcasts 0. +6 tests (incremental staging, restart, persistent-preseed, fresh-DB fallback); suite at 10-failure baseline. Co-authored-by: Matt Johnson <mj@k7zvx.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-05 15:18:48 -06:00
# The gate keys on the event's ``source`` (not the adapter name), and
# this adapter emits BOTH "firms" (raw hotspots) and "firms_fusion"
# (consolidated growth) — mark both seeded, exactly as a non-empty
# ingest would (store._ingest marks every source it touched).
feat(env): native adapters broadcast only newly-RECEIVED items (no backlog) (#43) Replace the native path's "scan accumulated state + suppress what we've already broadcast" model with "broadcast only what newly arrived from the API this poll." Storage is unchanged (self._events + firms_pixels etc. are populated for EVERY received item, so the LLM/get_active backlog is intact); only the BROADCAST decision changes. - env/store.py: per-adapter in-memory seen-set (_seen) + _seeded. First data-bearing poll for an adapter seeds keys and emits NOTHING (that batch is pre-existing backlog); later polls emit only keys not seen before. Restart => empty sets => next poll re-seeds silently. Structurally impossible to broadcast backlog on cold start / restart / re-enable. Key = external_id -> event_id -> content hash, namespaced per adapter. self._events[key]=evt still runs unconditionally (storage preserved). - Fixes the ~175 (roads511) / ~782 (wzdx) cold-start bursts AND the latent quake/nws version (they only looked safe because Central pre-populated their broadcast tables). - env/satpass.py: broadcast on AOS IMMINENCE (now < aos <= now+lead, broadcast_lead_seconds default 3600), future-only; window_hours still governs prediction depth. Strict norad_ids post-filter + fixed _parse_norad_ids char-iteration bug (cause of GOES/METEOR leak). - Central path + broadcast-state tables untouched (native-only gate). 13 new tests; full suite at 10-failure baseline (1697 passed). Co-authored-by: Matt Johnson <mj@k7zvx.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-05 14:38:40 -06:00
store._seen = {}
fix(env): leak-proof received-delta — durable persistent baseline + non-empty seed guard (#44) Two live backlog-broadcast leaks traced to the in-memory first-poll seed: (1) incremental-fetch adapters (wzdx: registry tick [0 events] then feeds tick [many]) got marked _seeded on the EMPTY first tick, so the real batch next tick all looked "new" and broadcast; (2) in-memory seed lost on restart. Fix — durable baseline + guard: - _seed_from_persistent() at store init: pre-load already-received item keys from the persistent hazard tables into self._seen, so nothing ever received can re-broadcast (immune to fetch staging + restart). Only sources whose native emit key PROVABLY equals a persistent key are durably seeded: wzdx (traffic_events.external_id) + usgs_quake (quake_events.event_id). Resilient (per-table try/except; missing table -> skip). - _seen_key() now namespaces by evt["source"] (matches persistent tables), via shared _key_ext/_key_eid helpers used by both seed and live emit so they can't drift. - non-empty-seed guard: _ingest marks only sources that carried >=1 event this poll as _seeded -> an empty first tick can never seed-then-leak. This is the root-cause fix; covers all adapters (roads511/traffic fetch atomically per tick, so the guard fully protects them). - storage untouched (self._events populated for every event); Central path/deciders untouched. Live-DB verified: seed pre-loads 784 wzdx + 8 quake keys -> a live wzdx poll of 784 known zones broadcasts 0. +6 tests (incremental staging, restart, persistent-preseed, fresh-DB fallback); suite at 10-failure baseline. Co-authored-by: Matt Johnson <mj@k7zvx.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-05 15:18:48 -06:00
store._seeded = {"firms", "firms_fusion"}
# Adapter-level cold-start seed is also a first-fetch silence gate; this
# steady-state test wants the growth wire, so mark the adapter seeded too
# (cold-start fusion suppression is covered in test_firms_cluster_f3).
a._firms_seeded = True
assert a.tick() is True
store._ingest("firms", a)
emitted = [c.args[0] for c in bus.emit.call_args_list]
cats = [e.category for e in emitted]
assert cats == ["wildfire_growth"], cats
assert emitted[0].summary.startswith("🔥 Pine Gulch")
def test_tick_with_no_fire_emits_nothing(self, monkeypatch):
# Pixels with no seeded fire: attribution misses, cluster dead ->
# zero fusion, and raw hotspots still suppressed.
_patch_fetch(monkeypatch, _csv([
{"lat": 43.5, "lon": -115.5, "t": "1300"},
{"lat": 43.6, "lon": -115.4, "t": "1305"},
]))
a = _adapter()
a.tick()
fusion = [e for e in a.get_events() if e.get("source") == "firms_fusion"]
assert fusion == []
chore(central-ripout 2d-ii): remove dead fire scraps (NOT the 'rewrite' — there wasn't one) (#166) * chore(fire-ripout 2dii-a): remove dead wildfire_growth cutover path + handle_firms wildfire_growth events are always fully precomposed (env/firms.py sets _meshai_precomposed=True + title=<wire from _render()>) and the category is not in cutover.NATIVE_ALWAYS_DECIDE, so the registered fire formatter could never be reached live via compose_mesh_message for this category -- the precomposed-title bypass always won first. Removes the dead is_cutover("wildfire_growth") NEW-PATH branch in fire_fusion._handle_pass_boundary (kept the live legacy leg that calls _render unconditionally) and the now-unreachable wildfire_growth formatter registration in notifications/formatters/__init__.py. Also removes handle_firms, the dead Central NATS-envelope entrypoint (zero live production callers -- Central's consumer that drove it is gone), plus its envelope-specific filtering helpers (_confidence_passes, _in_bbox, _coerce_severity, _log_event, FIRMS_CONFIDENCE_FLOOR/FRP_FLOOR/BBOX_OPTIONAL) that had no live callers left. The shared, LIVE fusion core (_ingest_pixel_core, attribution, clustering, growth/spotting/halt) and _parse_acq_epoch are unaffected -- still the engine behind the native ingest_hotspot_pixel entrypoint. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * chore(fire-ripout 2dii-b): remove dead handle_wfigs entrypoint handle_wfigs (the dead Central NATS-envelope entrypoint) had zero live production callers -- Central's consumer that drove it is gone. The LIVE WFIGS path is env/fires.py (native adapter) -> env/store.py::_emit_event, forced onto gating.fire.decide + the shared fire formatter via cutover.NATIVE_ALWAYS_DECIDE, independent of this dead handler. Removes handle_wfigs and its private-only helpers (_coerce_severity, _log_event, _log_event_returning_id) that had no callers left. _render remains -- it is called directly by env.fire_fusion._handle_pass_boundary on the FIRMS wildfire_growth path, and is used as a byte-identity oracle by tests against the shared, live fire formatter. _build_canonical, _attach_commit_handles, _fire_too_old_to_announce, _now, _cleanup_stale_fires and WFIGS_BROADCAST_COOLDOWN_S are kept -- each has its own direct test coverage independent of handle_wfigs. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * test(fire-ripout 2dii-c): rewire fire tests off the deleted dead entrypoints handle_firms and handle_wfigs (both removed in the prior two commits) were used throughout this test suite purely as convenient DRIVERS for live logic (pixel attribution/clustering/growth/spotting/halt fusion; the WFIGS New/Update/cooldown/closed decider). Per-file disposition: Rewired to drive the LIVE native entrypoint directly (no behavior change -- same underlying _ingest_pixel_core / gating.fire.decide engines): - test_fire_tracker_phase1/2/3.py: handle_firms -> ingest_hotspot_pixel - test_firms_refactor.py: handle_firms driver helpers -> ingest_hotspot_pixel; dropped the dead wildfire_growth cutover test + the fully-redundant TestNotCutoverLegacyVerbatim class (already covered by test_firms_native_fusion.py's TestIngestGrowth/Spotting/Halt against the live entrypoint); wildfire_growth formatter registration test now asserts None (matches source change). - test_wfigs_handler.py: handle_wfigs -> _render (the live WFIGS renderer) called directly, for the anchor-priority + missing-acres cases that exercise shared/live code (_location_anchor). - test_fire_refactor.py: handle_wfigs -> gating.fire.decide() driven directly, with state written the same unconditional shape the live native path uses; TestGateSequenceParity trimmed to focus on the tombstone/closed lifecycle step (not covered by test_fire_native_growth.py's native-adapter New/Update/cooldown coverage). - test_tombstone_broadcast.py, test_fire_age_gate.py: handle_wfigs -> gating.fire.decide() driven directly; same assertions, off the dead path. Deleted as pure dead-entrypoint contract testing with no live equivalent: - test_firms_native_fusion.py::TestCentralPathUnchanged (2 tests) -- existed only to guard handle_firms's own envelope parsing. - test_firms_handler.py: confidence/FRP/bbox filtering, missing-coords, non-firms-adapter guard, event_log accounting (all handle_firms-specific; the native adapter filters upstream in a different code path). Kept + rewired: the shared _ingest_pixel_core dedup behavior and _parse_acq_epoch's int/short acq_time parsing (both genuinely live/shared). - test_wfigs_handler.py: envelope field-extraction (acres-fallback chain, IA-placeholder-as-name), tombstone/perimeter subject -> event_log, New/Update/cooldown decision + audit-row wiring (all redundant with tests/test_fire_native_growth.py's coverage of gating.fire.decide() through the real native adapter). - test_tombstone_broadcast.py::test_commit_callback_flips_handled -- asserted handle_wfigs's own event_log-row flip on commit, a Central-only concept the native path never used. - test_tail_followups.py::test_wfigs_tombstone_stamps_column -- asserted handle_wfigs's own inline tombstoned_at UPDATE; no live producer emits _kind=wfigs_tombstone today, so nothing in the live system stamps it. Both live renderers stay covered: fire_format (wildfire_declared/incident, via test_fire_refactor.py + test_fire_native_growth.py, untouched) and _render (wildfire_growth via FIRMS, via test_fire_tracker_phase2.py + test_firms_native_fusion.py's TestIngestGrowth, both driving the live ingest_hotspot_pixel entrypoint). Collection: 2010 -> 1974 tests (net -36: dead-only tests deleted, live behavior rewired 1:1 or consolidated onto already-existing native-path coverage). Full suite: 1974 passed, 0 failed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Matt Johnson <mj@k7zvx.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-17 22:11:00 -06:00
# chore/ripout-2dii: section 3 ("Central-path guard — extraction did not
# change handle_firms") REMOVED. `handle_firms` (the dead Central
# NATS-envelope entrypoint it drove) has been deleted from
# `meshai.env.fire_fusion` -- zero live production callers, and this section
# existed purely to prove that entrypoint was unchanged by the extraction.
# The shared core it guarded (_ingest_pixel_core) is exercised directly by
# every test above via `ingest_hotspot_pixel` (the live native entrypoint).