0.1.15

Performance

• Dropped the per-decode objectToRefId reverse-lookup Map rebuild. The decoder already tags every Schema/ArraySchema/MapSchema instance with ~refId, and DataChange carries refId directly — both are used instead. Eliminates ~137 KB/op of garbage per decode on a 200-player state.
• Unified previousResultsByRefId / currentResultsByRefId into a single persistent cache plus a reused visitedThisPass set. Removes a per-render Map allocation and copy loop in useColyseusState.
• Memoized getSchemaFieldNames per Schema constructor (was Object.values(metadata).map(...) on every visit).
• Replaced Object.keys(prev).length !== Object.keys(next).length MapSchema size check with a count vs node.size comparison (no array allocations).
• Parent-chain dirty walk now short-circuits once it hits an already-dirty ancestor.

Measured on a synthetic "50 of 200 players mutate each tick" benchmark: ~25% faster, ~99% less allocation per tick.

Tests

• +60 tests covering: selectors, primitive ArraySchema / MapSchema, Schema reference reassignment, MapSchema.clear(), multiple hooks sharing a subscription, late-binding state/decoder, useRoomMessage, useLobbyRoom, useQueueRoom, createRoomContext, createLobbyContext, and coexistence with getDecoderStateCallbacks (onAdd / onRemove / onChange / listen).
• New opt-in microbenchmark at src/tests/bench.test.ts (BENCH=1 NODE_OPTIONS=--expose-gc npx vitest run src/tests/bench.test.ts).