Fix circuit breaker probe lockup, cache key collision, evaluators leak, and HOCON list corruption

[EtaCassiopeia]

Summary

CircuitBreakerProvider: half-open probe slot permanently stuck on application errors

Application errors (FlagNotFound, TypeMismatch) in half-open state did not call recordSuccess or recordFailure, leaving the probe slot locked. All subsequent requests were rejected forever. Now treats application errors as success for circuit purposes since the provider is reachable.

CachingProvider: cache key hash collision serves wrong flag values across users

CacheKey used a 32-bit contextHash: Int. Two different evaluation contexts could collide, causing user A to receive user B's cached flag value. Replaced with contextFingerprint: String using stable sorted serialization of context attributes.

CachingProvider: unbounded evaluators map leaks memory

The evaluators: ConcurrentHashMap side-channel had no eviction — entries accumulated for every unique cache key over the provider's lifetime. Changed evaluators.put() to evaluators.remove() in the Lookup callback so entries are cleaned up after cache population.

HoconProvider: LIST config values become list of nulls

configValueToSdkValue called .asObject() on each list element, which returns null for scalar values. A config like allowed-regions = ["us", "eu"] produced [null, null]. Removed the .asObject() call.

Event bridge: getOrThrowFiberFailure can crash Java SDK event thread

All event bridge handlers used getOrThrowFiberFailure() which could throw on defects (e.g., Hub shutdown), killing the Java SDK's internal event dispatch thread. Replaced with getOrElse to absorb unexpected failures.

Test assertions: overly broad result.isLeft checks

• BehaviorControlsSpec: setErrorMode(ProviderNotReady) now asserts the specific error type
• EvaluationTimeoutSpec: per-call timeout tests now assert ProviderError type