Does this cope with geofence reloads?

Thanks, that's a good point. I managed to forget we have reload on geofence - It wouldn't crash, but there was a race.

Fixed with last commit.

Along with other improvements and GC changes, CPU went down almost 3 times :)

Review feedback

1. nestTree should also use atomic.Value for consistency

statsTree was changed from *rtree.RTreeG to atomic.Value to support concurrent access during geofence reload. nestTree has the same access pattern but is still a bare pointer. Both should use atomic.Value or neither should.

Action: Change nestTree to atomic.Value and update MatchNestGeofence to load from it the same way MatchStatsGeofenceWithCell loads from statsTree.

2. Move build-phase mutex out of S2CellLookup struct

S2CellLookup.mu is only used during BuildS2LookupFromFeatures to protect concurrent writes from worker goroutines. After build, the map is immutable and Lookup() reads without locking. The mutex should not live in the struct.

Action: Remove mu sync.Mutex from S2CellLookup. Instead, use a local sync.Mutex inside BuildS2LookupFromFeatures and pass it to addArea/addEdgeCell (or make them unexported helper closures within the build function). This makes the immutability contract explicit — the struct returned from build has no synchronization because it needs none.

3. Config comment should mention memory cost

The 10MB geofence → 616MB lookup table is a 60x memory blowup. The config comment just says "Use S2 cell lookup before rtree" which doesn't hint at the tradeoff.

Action: Update the config comment and example to mention the memory cost, e.g.:

s2_cell_lookup = false      # Pre-compute S2 cell lookup for faster geofence matching. Trades memory (~60x geofence file size) for ~2-7x faster lookups. (default: false)

4. Test improvements

The test has //go:build ignore and requires a local geofence file, so it can never run in CI. Also, the comment "S2 lookup may return fewer areas for edge cells - this is expected" is misleading — the design correctly removes edge cells entirely so the rtree fallback always gives the full answer. Mismatches should always be 0.

Action:

• Change the mismatch comment to: "Mismatches should be 0 — edge cells are excluded from the S2 lookup, so the rtree fallback always provides the complete result."
• Assert mismatches == 0 with t.Errorf so the test actually fails if the invariant breaks.
• Consider adding a small unit test (without //go:build ignore) that constructs a synthetic geofence in-memory and verifies S2 lookup + fallback matches rtree for points inside, on edges, and outside.

(I asked claude to review in a form that you could feed straight to your AI :-)

@jfberry :) done