fix(engine): release worker buffers in coordinated close to avoid JVM crash on binding fault

[jfallows]

Problem

A hot-path binding exception (e.g. a null-exchange NPE in a client stream factory) is caught by EngineWorker.doWork and rethrown wrapped in AgentTerminationException, terminating the worker agent. Agrona's AgentRunner then runs the agent's onClose on the agent thread, which released the worker's memory-mapped buffers in advance. The engine-coordinated Engine.close() then calls drain(), reading that worker's already-unmapped streams ring buffer via Unsafe.getLongVolatile → SIGSEGV / JVM abort (confirmed via hs_err_pid fatal-error log).

The intended policy is preserved: an uncaught binding exception should stop the engine — but it must not also crash the JVM.

Fix

• EngineWorker — onClose no longer releases any memory-mapped resources. All releases (targets, streamsLayout, bufferPoolLayout, debitors, creditor, eventWriter) are deferred to doClose's finally, which runs on the engine thread after runner.close() has stopped the agent. A self-terminated worker therefore no longer unmaps buffers the engine — or peer workers (shared creditor/debitor budget buffers) — still reference.
• Engine.close — skips draining a worker whose agent already terminated (!worker.runner().isClosed()); such a worker will never consume again, so draining it would otherwise spin until the 30s timeout.

Validation

• Repro (NPE injected in a binding newStream): pre-fix → JVM abort (exit 134, hs_err_pid); post-fix → no crash, no hs_err, ~1.0s (no drain hang). The worker stops and the engine closes cleanly.
• ./mvnw verify -pl runtime/engine (unit + EngineIT + jacoco + checkstyle): green on the develop base.

🤖 Generated with Claude Code

[jfallows]

LGTM