feat(ingest): offline geo-resolution engine + gazetteers (PR1, no lens)

[k2jac9]

What

A reusable, 100%-offline geo-resolution engine that turns whatever geo a record carries — a coordinate, an intersection id/string, an address, or a ward/neighbourhood — into (lat, lng, quality) using only local committed gazetteers (no geocoding API, no CDN). Makes the "ingest any city's open data" claim real for the messy-geo datasets we currently skip (e.g. 311).

Scope: engine + committed gazetteers + tests. No consumer/lens wires it into the kernel, so it is behaviour-neutral by construction — the golden numbers (J $323,222 → $105,050) are byte-identical.

Engine — src/urbanos/risk/ingest/georesolve.py

• Cascade: measured → intersection-id → intersection-string → address → area-centroid → unresolved. First in-bbox hit wins; every coord (input and gazetteer-stored) is validated via in_toronto_bbox, so an out-of-region coord falls through instead of becoming a (0,0) pin.
• Confidence + fall-out: each tier emits confidence ∈ [0,1] (1.0 measured/exact, the difflib ratio for fuzzy, 0.30 for area); a candidate below max(min_confidence, tier_gate) falls out to the next tier.
• Built from existing primitives (normalize_address, in_toronto_bbox, _STREET_TYPES/_DIRECTIONS, _find_col/_clean_address/_coord/_read_rows) + stdlib difflib. normalize_address is untouched; the new canonical_street() helper reuses its constants. No new dependency.
• Offline-safe: resolve() never raises; load_gazetteers() is empty (not an error) when files are absent; process cache + reset_gazetteer_cache().

Gazetteers — scripts/fetch_gazetteers.py → demo_data/

Downtown Centreline intersections (~4000, 6-dp), city ward + neighbourhood centroids (pure-Python vertex mean). Offline-safe; raw GeoJSON never committed.

Validation — scripts/validate_gazetteers.py

Holdout harness re-resolves each intersection from its description alone, measures metre error vs truth (5 m tolerance, matching the tests), writes a fallout report capped at 60 rows (over-tolerance collisions first). Real run: 66.2% within 5 m, p50/p95 = 0 m. Reporting only — never fails the build.

Tests / verification

• tests/test_georesolve.py — 21 tests (cascade priority, accuracy, fuzzy bound, confidence/fall-out, bbox rejection, offline degrade, determinism, committed-slice guard, cache parity).
• Full suite: 684 passed, 1 skipped locally; golden numbers byte-identical.
• make demo-data regenerates the slices + report.