feat(titles): canonical incident names per adapter

[nullhack]

Problem

Dashboard incident titles were verbose and inconsistent — GDACS tropical cyclone titles were 100+ char run-on sentences (e.g. "Red notification for tropical cyclone BAVI-26. Population affected by Category 1 (120 km/h) wind speeds or higher is 36.778 million ."), USGS titles were "M 7.5 - 20 km ESE of Yumare, Venezuela", WHO titles varied wildly.

The dashboard's synth override (generate_dashboard_data.py:504) only fired when max_mag is not None, which failed for GDACS TC/Flood records with non-numeric severity prose — so the verbose raw name leaked straight through to canonical_name.

Fix

Standardised format produced at ingest by each adapter:

{Type} {Identifier} {Smallest-place}, {Country} {YYYY-MM-DD}

Diseases omit the Type word (disease name leads):

{Disease} {Place} {YYYY-MM-DD}

Samples (before → after)

incident_id	Before	After
20260624-VE-EQ	Earthquake 20 km ESE of Yumare, Venezuela 2026-06	Earthquake M7.5 Yumare, Venezuela 2026-06-24
20260701-CN-TC	Red notification for tropical cyclone BAVI-26. Population affected... (118 chars)	Tropical Cyclone BAVI-26 China 2026-07-01
20260606-CN-FL	Orange flood alert in China	Flood Hunan, China 2026-06-06
20251014-ID-VO	Eruption Kanlaon	Volcano Kanlaon, Philippines 2025-10-14
20250905-CD-DI	Ebola virus disease – Democratic Republic of the Congo	Ebola Kinshasa, DR Congo 2025-09-05
20250528-XX-DI	COVID-19 - Global Situation	COVID-19 Global 2025-05-28
20251221-XX-DR	Drought Green	Drought 2025-12-21

Architecture (adapter-centric, not a one-fits-all parser)

Each adapter owns _extract_canonical_name(raw_fields, places, report_date, incident_type) -> str, called at ingest so SourceReport.name is born canonical. Shared utilities in disaster_report/_title_format.py:

• format_title(*parts) — joins non-empty parts with single spaces
• smallest_place(places) -> (smallest, country) — locality > subdivision > country
• format_place(smallest, country) — "smallest, country" or just "country"
• normalise_subdivision(name) — strips accents + admin suffixes ( Sheng, SAR, Province, etc.)
• normalise_country(name) — short aliases (DR Congo, UK, etc.)

Per-adapter extraction

• USGS: M{mag:.1f} from raw_fields.mag + smallest_place.
• GDACS: dispatches by incident_type — EQ magnitude from numeric severity or regex on severitytext; TC storm name from eventname or title regex; VO volcano name from title (volcano name IS the locality); FL/WF/DR omit the identifier.
• WHO: disease short name via prefix lookup table (25 entries covering all 28 production incident_type values); place resolved from the title suffix (split on dash variants, scan_countries on the suffix) to avoid body-scan medevac noise — falls back to smallest_place(places) or "Global" when no places.

Dashboard simplification

Dropped the synth override at generate_dashboard_data.py:503-505. canonical_name now = inc["name"] directly.

Backfill migration

scripts/backfill_canonical_names.py: one-shot, per-adapter dispatch, targeted name: line replacement only (rest of YAML preserved byte-for-byte), --dry-run flag, idempotent. Data branch backfilled separately (164 reports: 67 USGS + 52 WHO + 45 GDACS).

Tests

52 new tests: 27 _title_format + 5 USGS + 10 GDACS + 10 WHO (4 new for title-suffix parsing).

Gates

• ruff check: clean
• pyright: 0 errors
• mypy.stubtest: clean (10 modules)
• pytest: full suite green

Related

• Data branch backfill: 0ded726 chore(data): backfill canonical report names (164 reports)
• Dashboard publish: follow-up (regenerate from backfilled data)