refactor: harden ThreadForge task lifecycle

[wuuJiawei]

Summary

• Isolate control timers from user scheduling work.
• Make task lifecycle, cancellation, physical termination, and scope close deterministic.
• Fix interruptible channels, completion-order fail-fast, await interruption, registration races, and context restoration.
• Make scheduler and DelayScheduler shutdown behavior explicit.
• Validate scheduling durations and retry multipliers.
• Extend CI coverage to SLF4J, Micrometer, examples, and benchmarks.

Verification

• Core: mvn -B -ntp clean verify and mvn -B -ntp clean install
• SLF4J integration: mvn -B -ntp -f integrations/threadforge-slf4j/pom.xml clean verify
• Micrometer integration: mvn -B -ntp -f integrations/threadforge-micrometer/pom.xml clean verify
• Examples and benchmarks package successfully.
• Core tests pass on JDK 8, 11, 17, and 21.

Please review; do not merge yet.

[wuuJiawei]

Verification update: concurrency and load testing

This PR has now been validated with repeated unit/integration tests, executable example scenarios, and sustained concurrent JMH load.

Unit and integration verification

Area	Executions	Result
Core full suite	617 tests x 4 runs = 2,468	PASS
SLF4J integration	4 tests x 10 runs = 40	PASS
Micrometer integration	1 test x 10 runs = 10	PASS
Lifecycle/race stress group	507 tests x 10 additional runs = 5,070	PASS
Example runtime scenarios	4 examples x 10 runs = 40	PASS

The lifecycle/race stress group covered:

• SchedulerIsolationTest: 100 tests per run
• TaskLifecycleTest: 107 tests per run
• ScopeRegistrationRaceTest: 300 tests per run

Total JUnit executions across this verification pass: 7,588, with 0 failures, 0 errors, and 0 skipped tests.

All four examples were launched as real JVM processes. Their stdout was checked exactly on every run, and every process exited normally.

Concurrent JMH load

Environment: JDK 17.0.8, 10 logical CPUs, 32 GB RAM.

Scenario	Concurrency	Average time	Approx. aggregate throughput
Fan-out / 8 child tasks	1	52.517 us/op	19.0k ops/s
Fan-out / 8 child tasks	8	332.335 us/op	24.1k ops/s
Fan-out / 8 child tasks	32	1.353 ms/op	23.7k ops/s
Fan-out / 8 child tasks	64	2.708 ms/op	23.6k ops/s
First-success / 3 candidates	1	34.485 us/op	29.0k ops/s
First-success / 3 candidates	8	138.155 us/op	57.9k ops/s
First-success / 3 candidates	32	566.802 us/op	56.5k ops/s
First-success / 3 candidates	64	1.143 ms/op	56.0k ops/s

The sustained 64-concurrency runs used 2 x 2-second warmups plus 5 x 5-second measurement windows:

• Fan-out: approximately 590k scope operations and 4.72M child-task submissions during measured windows.
• First-success: approximately 1.40M scope operations and 4.20M candidate-task submissions during measured windows.
• No exceptions, deadlocks, permanent hangs, or resource-exhaustion failures were observed.

Performance observation

The structured lifecycle has measurable microbenchmark overhead versus the plain JDK baselines:

• Fan-out: approximately 4-7x slower than ExecutorService/CompletableFuture baselines.
• First-success: approximately 3-4x slower than ExecutorService.invokeAny.
• Throughput on this 10-core host saturates around 8-32 concurrent benchmark workers.

This is not enough to call the change a performance regression because there is no same-host pre-fix baseline. It is recorded here as an optimization target and not hidden as a correctness failure.

Validation boundary

This is production-shaped local concurrency/load testing of the ThreadForge library, not replay of real production traffic. The repository does not contain a deployable HTTP service, production request model, or production endpoint.

Remote CI remains green on JDK 8, 11, 17, and 21, including integrations, examples, and benchmarks. git diff --check also passes.