Cut Google Maps spend: reverse-geocode cache, Details mask trim, confirm skip, closure persistence

[timwangmusic]

The no-wire-change batch from the 2026-08-01 geo cost/freshness audit. Four cost cuts plus one zero-cost freshness fix:

1. Reverse-geocode cache (the big one)

Every nearby scan buys a reverse-geocode via processLocation — and on a warm place cache it was the only Google call the scan made, so it was ~100% of warm-scan spend. Results are now cached per ~8 km search cell (geocode:reverse:<cell>, 30-day expiry per Geocoding caching terms). The RedisClient.ReverseGeocode stub is now a real implementation; processLocation, PoiSearcher.ReverseGeocode, and the /v1/reverse-geocoding handler all route through it.

2. Details field-mask trim

detailed_search_fields drops name and user_ratings_total: no Details consumer reads either (both arrive free with every Nearby/Text Search result), and user_ratings_total alone pulls every Details call into the Atmosphere billing tier. The AddUserRatingsTotal migration passes its own single-field list and is unaffected. editorial_summary stays (trip planner reads it) — note that means the Atmosphere tier still applies until that's revisited. Pinned by TestDetailedSearchFieldsMask.

3. Confirm skips redundant Details

/v1/place-search/confirm bought a full max-tier Place Details call even when the cached record's details were current — data restoreCachedDetails restores anyway. It now applies the same placeDetailsAreCurrent rule the nearby path already trusts. This also makes Offerbee's wallet-change re-confirm of manual rows free in the common case.

4. Photos-only mask in stale-photo recovery

photos_client.go requested the entire Details mask to recover one photo reference; now ["photos"].

5. Closures are persisted (freshness fix, zero cost)

Cold searches filtered non-OPERATIONAL results before the cache write, discarding the closure signal entirely — a permanently closed place kept its OPERATIONAL record and cache membership forever. The write now includes closures (new persistAndFilterSearchResults owns write + response filter); the response-side behavior is unchanged, and the existing read-side Operational filter retires closed places from cache serves.

Testing

• TDD throughout: new suites reverse_geocode_cache_test.go (round-trip, cell scoping, expiry, cache-first searcher + processLocation), add_searched_place_currency_test.go (skip-when-current, buy-when-stale, buy-when-uncached), closure_persistence_test.go (persist + response exclusion + read-path retirement), config_yaml_test.go (mask pin).
• go vet ./... clean; full go test ./... green (7 packages).
• The photos one-liner has no unit test — that call site has no stub seam; covered by build + the narrowed literal.

Deliberately not in this PR: hoursKnown wire flag (API change), text-search caching + timeout salvage (larger), MapsLastSearchTime dead-field purge (Redis ops).

🤖 Generated with Claude Code

https://claude.ai/code/session_01DjdGnZM6cUBeg6jWwoQfU3

[timwangmusic]

Added a fail-safe flag on top of the closure-persistence change: PlaceSearchRequest.BusinessStatus is replaced by IncludeClosedPlaces (zero value = non-Operational places always filtered, from cache reads and cold-search responses alike). Previously a caller had to remember BusinessStatus: POI.Operational to get filtering — with closures now actually present in the cache, a forgotten field would have served closed places to planners. All three production call sites already requested filtering, so behavior is unchanged; only the failure mode is gone.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DjdGnZM6cUBeg6jWwoQfU3

[tim-eternos]

lgtm