feat: add Micrometer observability to GlueSync listener and Kafka receiver

[jamespfaulkner]

Summary

Adds structured JMX metrics to ApiaryGlueSync so every HMS event outcome is observable without log-scraping.

• Structured event counters — recordEvent(operation, result, outcome) emits a listener.event counter tagged with operation (e.g. create_table), result
  (success/failure/ignored), and outcome (e.g. OperationTimeoutException, created, other). All failure paths, success paths, and ignored events are covered.
• JMX tag promotion — TaggedObjectNameFactory + taggedNameMapper() encode Micrometer tags as proper JMX ObjectName key properties
  (metrics:name=listener.event,operation=create_table,result=failure,outcome=OperationTimeoutException), making them filterable in JConsole/JMX tooling without string-parsing.
• Bounded cardinality — ~8 operations × 3 results × 3–5 outcomes gives ≤72 unique counters.

Bug Fixes

• Shade-safe outcome classification — toOutcome() was silently returning "other" for all AWS errors. The Maven shade plugin relocates com.amazonaws.* to
  com.expediagroup.apiary.shaded.com.amazonaws.* at runtime, so Class.isInstance() against the original class literals never matched. Fixed by walking the exception class
  hierarchy using getSimpleName(), which is package-agnostic and survives relocation.
• Duplicate JmxMeterRegistry prevention — configuredRegistry() is now synchronized and guards against double-registration under concurrent HMS init.
• consumer.close() guarantee in KafkaMessageReader — consumer.close() was not called if kafkaClientMetrics.close() threw; fixed with a try-finally.

Performance

Registration cost is one-time per unique (operation, result, outcome) combination. After the first call, computeIfAbsent returns the cached Counter and the hot path is
counter.increment() — a lock-free LongAdder.add(1). TaggedObjectNameFactory never touches the request thread again.

recordEvent() is called inside the per-partition loop on onAddPartition/onDropPartition, where a single HMS event can carry hundreds of partitions. The string concat
forming the cache key allocates a small object per call, but computeIfAbsent hits the already-populated ConcurrentHashMap after the first call (hash lookup only). This is
negligible compared to the synchronous Glue API call that precedes each partition operation. All ≤72 counters are registered within the first minutes of steady-state traffic.

Test coverage

Baseline test coverage was established in #136. This PR adds recordEvent verifications to all existing HMS event handler tests, plus:

• Failure metric tests covering OperationTimeoutException, InvalidInputException
• toOutcome hierarchy-walk tests: AWS subclass not in known list → "AmazonServiceException"; unknown exception → "other"
• TaggedObjectNameFactory unit tests
• Rename failure metering, partition-not-found path

---

🤖 Generated with Claude Code