feat(meteoalarm): merge MeteoFrance forecast days into one episode

[seevee]

Closes #37. Follows #68, which removed the green markers that would otherwise collide inside this merge.

Problem

MeteoFrance publishes one warning per calendar day, each running roughly 00:00 → 00:00 local, and the next day's bulletin goes live alongside the current day's for most of the day — live sampling found the overlap in 170 of 227 samples, spread across every hour rather than confined to an afternoon window.

The MeteoFrance content key from #41 includes a forecast-day component, so one multi-day heat or storm episode became one entity per day, with the id rolling over at midnight. That is what the reporter sees: two Vigilance jaune canicule entities, contiguous at the midnight boundary, and a fresh entity every morning.

Fix

Collapse a run of consecutive forecast days into one alert, keyed without the day component.

Region-picker mode explodes the bulletin first, into one alert per configured department, using the <area> blocks that already carry one areaDesc and one NUTS3 code each. The key becomes (sender, phenomenon, one department).

That choice is the crux, and it is measured rather than assumed. The department set a bulletin covers moves overnight:

awareness_type	08-03	08-04	delta
12	3	3	identical
13	1	1	identical
3 (thunderstorm)	83	54	+5 / −34
5 (heat)	57	54	+0 / −3

So a set-derived key — or its intersection with a multi-department config — splits the episode anyway. Per-department scoping is stable, and is safe because live sampling shows at most one live warning per (department, phenomenon, day) across 203 samples. It also replaces an area_desc listing up to 83 departments with the one the user selected.

The most severe day supplies the content wholesale, tie-broken to the earliest onset; the window widens to span the run. Blending fields would let the record contradict itself, since severity_normalized derives from awareness_level and the icon from event — a blend could read "Vigilance jaune canicule" while carrying an orange level. Per-day truth moves to a new episode_days attribute.

Result on live data

Replayed against a real France feed at a fixed instant:

entities per department BEFORE: {1: 22, 2: 6, 3: 20, 4: 39, 6: 1}
entities per department AFTER:  {1: 28, 2: 59, 3: 1}
total across 88 departments: 256 -> 149

FR211 (Ardennes), after:
  'Vigilance jaune canicule'  2026-08-03T16:04 -> 2026-08-05T00:00  days=['08-03','08-04']
  'Vigilance jaune orages'    2026-08-03T16:04 -> 2026-08-05T00:00  days=['08-03','08-04']

Nothing is lost in the process: 88 → 88 departments covered, 149 → 149 (department, phenomenon) cells, and zero departments losing a phenomenon in region mode. I checked this explicitly because a wholesale dominant-day copy could plausibly have dropped coverage.

Two things the plan got wrong, found while building

• Gap-day runs collided. Two live runs in one group both got the day-free id, so the alert store would have silently dropped one — the exact failure the merge exists to avoid. Second and later runs now re-add their first day to the key, which churns only the pending entity and never the one in effect. The case has never been observed live (0/227 samples), but it must not lose a warning if it happens.
• episode_days on single-day runs was pure noise restating the alert's own fields. Now empty unless something actually merged, keeping attributes sparse.

Scope and limitations

• Country-wide mode keeps the full-set key and still splits an episode when the footprint moves. Documented rather than silent: exploding per department there would turn France into roughly 150 entities.
• A gap of more than one calendar day starts a new run, read as a separate episode. That reading is unproven — 0 of 227 samples show a gap — and getting it wrong costs an extra entity rather than a lost warning.
• GPS mode is unreachable for MeteoFrance: the France feed carries zero geometry (0 polygons across 6742 area blocks), so _filter_by_polygon fails loud first. The branch is defensive only.
• A horizon/outlook filter was rejected — the merge subsumes it. Live depth is two forecast days, not the four the J/J+1/J+2/J+3 framing implies; that larger figure counted superseded messages.

Breaking change

MeteoFrance entity ids change once on upgrade; stale entities are safe to delete. Multi-department configurations also get a granularity change — one entity per configured department instead of one per bulletin — so they will see more entities than before, each stably named and identified. Both need announcing in the release notes and on #37.

Every other authority and provider is untouched, guarded by pass-through tests.

Also in this PR

• _filter_by_regions returns to pure filtering; identity is owned by the merge that runs after it.
• normalize exports SEVERITY_RANK and meteoalarm_awareness_severity so the provider ranks days on the same ladder normalization applies, instead of a second copy that can drift.

Verification

Check	Result
pytest tests -q	538 passed (was 521)
ruff check / ruff format --diff	passed / 58 files already formatted
mypy custom_components/cap_alerts	no issues, 22 source files

17 new tests in tests/test_meteoalarm_episodes.py. Payloads are built rather than stored as JSON fixtures — the shapes differ from one another by a field or two, so a builder keeps the difference legible where five near-identical fixture files would bury it. Every test injects now, since the merge is now clock-dependent.

Two existing tests changed, both anticipated: test_region_scope_stable_across_set_churn became test_region_filter_is_pure_filtering (its scope-stability assertion is now covered end-to-end by test_episode_stable_across_footprint_churn), and test_region_picker_intersects_nuts3 pins now inside its fixture's window.

Refs #37

Assisted-by: Claude:claude-opus-5