fix(tests): stop SessionLogActor diagnostic test racing the FlushTick

[Aaronontheweb]

Problem

SessionLogActorTests.Dispatcher_writes_audit_and_diagnostic_to_same_session_log is flaky on Windows CI. It failed on run 31715544081 (job 94499226187, 2026-08-13): after the full ~3s AwaitAssertAsync budget, the file contained the assistant line but not the diagnostic line, even though the write was correct.

Root cause (confirmed by dotnet-concurrency-specialist + akka-net-specialist):

• The test sent the durable audit TextOutput first (flushed immediately via WriteDurable), then the buffered SessionLogDiagnostic (Write only — no flush).
• Buffered diagnostics only reach the OS at a flush point: the 1s FlushTick (real wall-clock scheduler, not the test's FakeTimeProvider), the 256-write threshold, a later durable write, or PostStop.
• On a loaded 2-core Windows runner the tick can land past the 3s assert budget → flake. The test predates the batching refactor (refactor(logging): partition the log stream by session — session.log per session, daemon.log sparse, OTEL the union (#1472) #1499) that introduced the buffered diagnostic path.

Fix (test-only, deterministic)

Send the buffered diagnostic first, then the durable audit line — the audit write's WriteDurable → Flush() drains the buffered diagnostic synchronously in the same mailbox turn. This is the exact behavior documented in SessionLogActor.cs: "audit lines are natural flush points."

No production change: the buffered-diagnostic path is deliberate (batching, per #1499) and bounded in production.

Verification

• 10/10 stress runs pass in 433–702ms (previously ~1s every run, gated on the tick).
• Full SessionLogActorTests class: 4/4 pass.
• Fix removes the wall-clock dependency entirely, so the loaded-runner failure mode cannot resurface.

Notes

• Test-only change; no OpenSpec, no system-skill change.