fix(tests): deterministically gate two Windows-flaky actor cold-start tests

[Aaronontheweb]

Problem

Two actor tests flake only on the Windows CI runner (Ubuntu + macOS stay green in the same job). They surfaced most recently on the Test-windows-latest job of #1539, but neither test's file — nor any actor source — is touched by that PR, so these are pre-existing flakes that can red-X any PR:

• DiscordSessionBindingContractTests.Approval_response_sends_feedback
  → Assert.Contains() Failure: Filter not matched in collection after AwaitAssert failed, timeout [00:00:03] is over after [2] attempts
• SessionMemoryObserverActorTests.DistillMemories_only_persists_accepted_proposals_for_future_dedup
  → Timeout 00:00:05 while waiting for a message of type SessionDistillationCompleted

Root cause (same class in both)

A fixed short timeout racing a persistent-actor cold start under ThreadPool starvation. The assertions are correct; the deadlines were sized for the work, not the worst-case scheduling latency on a saturated Windows runner. The 2 attempts in 3.66s detail is the tell — the poll loop's retry continuation itself was starved, so it wasn't "the work took 3s," it was "the harness couldn't run."

• Binding test: DiscordSessionBindingActor is a ReceivePersistentActor. Output only renders after recovery → InitializePipeline → hydration → Active → UnstashAll; every OutputReceived before Active is stashed. The first AwaitAssertAsync (default 3s, from akka.test.default-timeout) had no readiness gate, so it polled GetPostedTexts() while that entire cold start was still in flight. CI logged "session pipeline initialized" ~3 ms after the poll gave up.
• Observer test: SessionMemoryObserverActor is a ReceivePersistentActor whose command handlers stash until RecoveryCompleted. It's an early persistent actor, so the in-memory journal/snapshot plugin cold start + recovery consumed the 5s ExpectMsg budget before the (fast) FakeChatClient distillation ran. CI logged session_observer_recovery_complete right at the deadline.

This is not a production defect — stash-until-RecoveryCompleted is the correct Akka.Persistence synchronization the parent legitimately relies on. It's a test cold-start-budget artifact, so the fix lives in the tests.

Fix — gate on real readiness signals (no widened timeouts, no sleeps)

• Binding test: await pipeline.Created.WaitAsync(ct) before the poll — a linear await on the signal RecordingSessionPipeline already exposes (completes inside CreateAsync, once recovery + init are done). This is byte-for-byte the pattern the Reminder_delivery_* and Stashes_messages_during_init siblings already use. The remaining tail is in-process, millisecond-scale, and the 3s poll now covers only that.
• Observer test: Ask an empty RecordAcceptedDistillationProposals([]) first. That command is answered immediately post-recovery with no Persist and no state mutation (empty-list early return), so it's a side-effect-free readiness ack; the generous 30s ceiling absorbs cold start without polling and the behavioral 5s windows are then measured from a warm actor. Mirrors the existing ReminderManagerActorTests generous-Ask precedent and CLAUDE.md's "Ask<Ack> so callers know a state transition occurred" rule.

Neither fix introduces Thread.Sleep/Task.Delay; both block on a real signal.

Validation

• dotnet build src/Netclaw.Actors.Tests — clean (0 warnings/errors)
• The two named tests — pass
• Full DiscordSessionBindingContractTests + SessionMemoryObserverActorTests classes — 83/83 pass
• dotnet slopwatch analyze — 0 issues
• ./scripts/Add-FileHeaders.ps1 -Verify — all headers present

Scope / follow-up

This fixes exactly the two tests that failed. Two independent specialist passes (Akka lifecycle + .NET concurrency) flagged that ~40 sibling tests across SessionBindingContractTests, DiscordSessionBindingContractTests, and SessionMemoryObserverActorTests share the same anti-pattern (a fixed poll/ExpectMsg racing first-actor cold start), and that DistillMemories_records_accepted_proposals_for_recovery_after_ack is a near-twin of the observer flake. I kept this PR tightly scoped to the observed failures rather than rewriting the family. If the maintainers want suite-wide insurance, the low-risk lever is a class-level akka.test.timefactor on those base test classes (the concurrency pass's recommendation) — happy to open a separate PR for that.