fix: pre-warm thread pool in CircuitBreakerSpecBase to fix cold-start latch flakiness

[Aaronontheweb]

Problem

Several async CircuitBreakerSpec tests fail intermittently on CI with a bare:

Assert.True() Failure
Expected: True
Actual:   False

Most recently on the Microsoft.NET.Test.Sdk 17.9.0 → 18.7.0 bump (#8282) — e.g. Must_reopen_on_exception_in_call (CircuitBreakerSpec.cs:283, HalfOpen latch) and Must_throw_exceptions_when_called_before_reset_timeout (line 311, Open latch). The Assert.True(false) is a CountdownEvent latch wait (latch.Wait(2s)) returning false.

Root cause

CircuitBreaker signals the test latches from its transition listeners, which run fire-and-forget on the thread pool (AtomicState.Enter → NotifyTransitionListeners → Task.Factory.StartNew, deliberately un-awaited so listeners never block state transitions). The half-open transition additionally depends on the HashedWheelTimer reset firing — also pool-driven. Meanwhile the test thread blocks in CountdownEvent.Wait.

The .NET thread pool injects new workers above the minimum at only ~1 per 500ms. On a cold/contended CI agent at process start, the fire-and-forget breaker call, the transition-listener signal, and the timer continuation can be starved behind that ramp and miss the fixed 2s latch budget. This is environmental timing, not a product bug — the SDK bump simply changed the test-host's startup footprint enough to expose a pre-existing latent race. (Confirmed: akka.test.timefactor is 1.0 on CI, so Dilated(...) is a no-op and the latch budget is a hard 2s.)

Fix

Pre-warm the thread pool in CircuitBreakerSpecBase so those work items get a worker immediately instead of waiting on throttled injection:

ThreadPool.GetMinThreads(out var minWorker, out var minIo);
ThreadPool.SetMinThreads(Math.Max(minWorker, Environment.ProcessorCount * 2), minIo);

This is the same remedy already used in the repo for the identical cold-start root cause (InMemoryPersistenceSpecConfig.EnsureThreadPoolWarmed()). Math.Max guarantees a previously-raised minimum is never lowered.

Test-only change — no public API touched (notably TestBreaker and its CountdownEvent latches are unchanged, per extend-only design) and no timeout values changed.

Verification

• Builds clean (Release, 0 warnings / 0 errors).
• Full Akka.Tests.Pattern suite green (Total: 71, Failed: 0).
• The two previously-failing tests pass 25/25 in a tight loop.

Local runs can't reproduce CI's cold-start starvation (the dev-box pool warms instantly), so local results confirm no regression; the fix is grounded in the root-cause analysis and the proven in-repo precedent.

Notes

Second of two flaky-test fixes split out from the #8282 investigation; the multi-node NodeChurnSpec fix is #8286. The broader observation that akka.test.timefactor is never raised on CI (making every Dilated(...) a no-op) is left as a separate, optional CI-config change.

[Aaronontheweb]

Pretty gross code fix.