Harden the known wall-clock test flakes

[lahma]

Addresses the recurring CI/local test flakes (e.g. the Windows CI failure on #2542). Five flake families were investigated to root cause; all share one pattern: wall-clock deadlines racing starved CPUs on loaded runners.

Test262 slow files (RegExp-*-escape-BMP, decodeURI byte-range loops)

These do seconds of legitimate work (64K-code-point regex loops; ~1.3M decodeURI calls) and pass in ~1–5 s isolated, but blow the 30-second wall-clock TimeoutInterval when the parallel suite starves the CPU — measured locally: 4 failures at exactly 30 s in a full run, 4 passes in 5 s isolated. A known-slow file list now gets a 5-minute budget; every other test keeps the tight runaway-script limit, so hang detection is not weakened.

AsyncTests task-interop / async-wake family

The CI failure above is EvaluateAsyncWithTaskInteropShouldWork: "Promise was rejected with value Timeout of 00:00:10 reached" — a Task.Delay(100) continuation queued past the default 10 s PromiseTimeout under thread-pool starvation on a two-core runner. (The event-loop wake path itself checks out: registration is double-checked and wake is broadcast; this is starvation, not a lost-wakeup bug.) The family now uses a generous 2-minute budget — these tests assert behavior, not latency. Tests asserting a timeout/cancellation fires keep tight budgets (starvation only helps them fire).

Bonus: ShouldRespectCustomProvidedTimeoutWhenUnwrapping — skipped as "Flaky test" — is un-skipped: it raced a 1 ms timeout against a 100 ms delayed task (under load the task could win and no exception was thrown); a never-resolving promise makes it deterministic.

TimeSystemTests.CanUseTimeProvider

Asserted the script''s new Date() within a fixed 100 ms window captured before a cold engine construction — regularly exceeded under load (the known net472 flake). Now bracketed between wall-clock readings taken around the evaluation; the frozen FakeTimeProvider half asserts exact equality.

Atomics.waitAsync early timeout (engine fix)

Surfaced during gating: no-spurious-wakeup-on-store.js failed because Task.Delay can complete slightly ahead of the requested interval (timer granularity/coalescing), letting the waiter resolve "timed-out" before the timeout actually elapsed — script-observable, and test262 asserts lapse >= timeout. The timeout timer now re-delays until a monotonic clock confirms the interval has fully elapsed — the same discipline the sync Atomics.wait loop already had. This complements the existing NonParallelFeatures serialization of the Atomics suites.

Gate

Jint.Tests / PublicInterface / CommonScripts green (including the un-skipped test on both TFMs), plus a fully green 99,429-test parallel Test262 run — previously the BMP quartet failed exactly this kind of run on this machine.

🤖 Generated with Claude Code