overture_canary: field-coverage regression gate (#416)

[chuofringer]

Summary

• Extends scripts/overture_canary.py (Stay current with the latest Overture release automatically (rollover, staleness alarm, schema canary) #219) with a third check: field-coverage regression, alongside the existing pin-staleness and per-theme column-presence checks (unchanged).
• Motivation (north-star): Widespread drop in brand field coverage for Brazil between releases ~90% of a random sample of well-known chains lost matched POIs (median -21%, some down to zero OvertureMaps/data#546 documented a release where a whole country's brand-matched POIs collapsed (median -21%, some chains to zero) with no schema change — the existing column-presence probe would have sailed straight through it. This closes that blind spot before the next pin bump (docs/PIN.md) adopts a damaged release blindly, protecting the honesty pillar.
• Probe set: 5 dense metro bboxes spread across continents — Central Paris (Europe), Manhattan (North America), Tokyo/Shibuya (Asia), São Paulo/Sé (South America), Lagos Island (Africa — the continent #546 was actually about). Each ~0.1° class (PROBE_RADIUS_M = 5_500), built with geo.bbox_around/geo.bbox_filter_sql, same helpers the runtime uses.
• Metrics per bbox, pinned vs newest release: places row count, addresses row count, and non-null rate of brand, confidence, and taxonomy.primary (the column overture._place_select_exprs actually reads for "category" — not basic_category).
• REGRESSION_TOLERANCE = 0.20 — chosen just under #546's documented median -21% drop, so the gate catches an incident of that shape without chasing ordinary release-to-release noise.
• MIN_ROWS = 50 floors row-count-backed metrics on the pinned side: a tiny denominator turns "one row missing" into a meaningless "100% drop".
• Reuses the existing markdown-report + exit-1 channel unchanged — the workflow's issue-creation path needs no changes and none were made (no .github/workflows/* touched).
• Gate is skipped (noted, not failed) when the pinned release already is newest, or when a probe can't read a release.
• Pure logic (compare_bbox_metrics, render_coverage_report) is separate from the network probe (probe_bbox_metrics, never run from pytest — same rule the script already states) and is covered offline with synthetic numbers in tests/test_overture_canary.py.

Test plan

• uv run pytest -q — 2375 passed (2368 baseline on main + 7 new offline tests), including existing column-check/pin-staleness tests unchanged
• uv run ruff check . — clean
• No .github/workflows/* files changed
• No new dependencies
• Offline coverage: healthy releases → clean; synthetic brand-collapse and row-count drops → flagged with correct table numbers; small-denominator metric → skipped, not flagged; pinned == newest → whole gate skipped with a note
• CHANGELOG.md ### Added entry added
• No schema/tool/server-surface changes → no benchmark regen needed

🤖 Generated with Claude Code

[chuofringer]

Lead review (loop)

Re-ran locally: 2375 passed (+7), ruff clean; verified zero .github/ changes (script-only per the issue's hard-rule scope).

Verified in review:

• Pure comparison/threshold/report logic is cleanly separated from the network probes (which correctly stay untested, per the script's own never-from-pytest rule).
• The 0.20 tolerance is sized against the documented incident (#546's median −21%), MIN_ROWS floors keep tiny denominators from producing noise flags, and skip conditions append to the clean-line suffix rather than findings — so the gate can never cause a false workflow issue.
• Probe set is continent-spread with Lagos deliberately covering #546's failure geography; taxonomy.primary confirmed as the column the runtime actually reads.
• Existing pin-staleness + column-drift behavior untouched on the normal path.

Acceptance criteria of #416 met. Ready for owner merge (loop does not merge).