[yanghang8612]

What does this PR do?

This PR is a first-step optimization toward #6697, focused on parallel execution in BatchValidateSign and VerifyTransferProof.

Share deadline-aware waiting and result collection, and cancel unfinished tasks on timeout, interruption, worker failure, or partial submission failure. Result reads are bounded when the VM's transaction-end CPU check applies.

Why are these changes required?

VerifyTransferProof previously used unbounded Future.get() calls after its timed latch wait. Even when the latch reaches zero, a worker may not have published its result yet. Unsuccessful exits could also leave submitted tasks running or queued.

This PR has been tested by:

Validated on JDK 17 / macOS ARM64 after rebasing:

• ./gradlew clean build -x test passed.
• The CI Checkstyle command, the Math usage script, and git diff --check passed.
• The framework suite completed 4,336 tests with zero failures (12 skipped), including all 24 terminal precompile VM scenarios and all 11 RelayServiceTest cases.
• All 15 parallel waiting/result collection tests passed.

The local full-suite run encountered two DNS environment failures in p2p tests: the system resolves reserved .invalid names to 198.18.* addresses. Those tests are unchanged by this PR; the CI matrix provides verification on its normal resolver.

Follow up

The broader work in #6697 remains open, including common precompile entry/exit checks, consistent timeout propagation, and internal checkpoints. Full-VM differential tests comparing receipts, energy accounting, state, and logs, along with replay compatibility and activation assessment, are also follow-up work.

Extra details

The branch is rebased onto release_v4.8.3 through 409a33f21e, including the upstream HELLO test fixture fix from #7060.

Preserves historical timeout behavior for debug, solidity, and pre-energy-adjustment execution. Precompile addresses, ABI, and energy pricing are unchanged. Cancellation of running work remains best effort.

[sunny-tron]

A couple of notes after reading through the timeout paths:

1. awaitParallel (via getCPUTimeLeftInNanoSecond()), readFuture and allowTimedOutResult each read getVmShouldEndInUs() separately. The equivalence with Program#checkCPUTimeLimit relies on all three using the same deadline as the calling program, so reading it once (for example at the start of collectParallel) and passing the value down would make that explicit and keep the three checks consistent.

2. The new tests exercise the helpers with hand-set deadlines. Since the change is meant to leave transaction results unchanged, a test at the Program/VMActuator level showing that a timed-out BatchValidateSign / VerifyTransferProof call still ends as OUT_OF_TIME with the same energy usage under energy adjustment would cover the part that matters most for compatibility.

[yanghang8612]

Thanks @sunny-tron. Both points are addressed in 9be9f29:

1. collectParallel now captures getVmShouldEndInUs() once and passes that deadline to the latch wait, future reads and timeout-result check. The historical latch deadline and the microsecond boundary used by Program#checkCPUTimeLimit are preserved. A focused helper test also verifies that collection reads the deadline only once.

2. Added 24 parameterized regression cases that run a terminal precompile CALL through the real Program, VMActuator.execute() and runtime result mapping for both BatchValidateSign and VerifyTransferProof. They cover normal completion, latch timeout and delayed Future publication across normal, energy-adjustment-disabled, debug and solidity modes. With energy adjustment enabled in normal mode, timeout cases assert OUT_OF_TIME, the transaction-end CPU check and full transaction energy consumption. The other cases check the expected return bytes and energy usage.

The tests control task completion through a scoped FutureTask construction mock, so no worker fields or executor replacement were added to production code. Successful cases have a 5-second budget; timeout cases retain a separate 150ms budget.

Validation: all 45 targeted tests passed with test retries disabled, and Checkstyle passed. An additional temporary 300ms scheduling-pause probe passed all 8 normal-completion cases; the probe was removed before committing.

[CodeNinjaEvan]

LGTM