perf(scheduler): atomically cache scheduler capabilities

[Arbousier1]

Summary

• supersede perf(scheduler): cache scheduler-capability lookups by target identity #108 with one immutable volatile capability entry per global, region, and entity scheduler cache
• retain identity-based hits and Paper null-capability negative caching while preventing a target from being paired with another target's scheduler under concurrency
• cover cache hits, negative caching, identity invalidation, and concurrent entity routing

Performance

Using the unchanged protected ServerSchedulerReflectionBenchmark locally on JDK 25:

• Folia: 23.889 ns/op -> 13.729 ns/op (about 42.5% faster)
• Paper: 19.857 ns/op -> 5.192 ns/op (about 73.9% faster)
• normalized allocation: Folia 67.429 -> 16.004 B/op; Paper 19.810 -> 16.000 B/op

The base-owned eight-shard scheduler-reflection gate remains authoritative.

Test plan

• ./gradlew test --tests top.ellan.mahjong.runtime.ServerSchedulerReflectionCacheTest spotlessCheck -PmahjongJavaToolchain=25 -PmahjongJavaTarget=25
• ./gradlew check assemble -PmahjongJavaToolchain=25 -PmahjongJavaTarget=25
• build base and candidate JMH jars with Java target 17 and run the protected scheduler benchmark unchanged
• base-owned GitHub performance A/B gate reports PASS_OPTIMIZED

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.