fix(test): make LlmSession Windows CI flakes deterministic

[Aaronontheweb]

Problem

Two LlmSessionIntegrationTests flaked on Windows CI (run 31435126237, PR #1859):

1. Reminder_redelivery_is_deduped_while_first_turn_is_in_flight — Assert.Equal(1, _fakeChatClient.CallCount) got Actual: 0 inside a 3s AwaitAssertAsync poll.
2. JoinSession_receives_SessionJoined_acknowledgement — AskTimeoutException : Timeout after 3.00 seconds.

Both are real-time-budget races on a loaded runner — same family as the #1858 Discord fix. This PR fixes them deterministically, without touching any timeouts.

Fixes

1. Reminder dedup test — event barrier instead of polling.

The CommandAck fires before the first turn's LLM call is scheduled (recall resolution + working-context mailbox hop run first), so polling CallCount == 1 raced the very thing it measured. Added a FirstCallEntered TaskCompletionSource to FakeChatClient, set when the first GetResponseAsync entry increments CallCount (before the gate). The test now awaits the signal — proving the turn is genuinely in flight (blocked on the gate) — then asserts CallCount == 1 directly. Event-driven, not wall-clock.

2. JoinSession test — warm up the cold spawn outside the timed Ask.

The first JoinSession cold-spawns the session actor through DI (fresh SQLite store, persistence recovery, serialization verification) inline in the parent mailbox — unbounded cost on a loaded runner that exceeds the 3s Ask. The test now warms the session first (Tell + probe ExpectMsgAsync<SessionJoined>), so the assertion Ask measures only the hot mailbox path. TurnCount == 0 / Title == null still hold — the warmup runs no turn.

The 30s ceilings are hang guards only, not pass/fail mechanisms — the pass/fail is event-driven (TCS completion / probe message receipt).

Verification

• Both tests pass; full LlmSessionIntegrationTests class 34/34 passes on Linux.
• 5x stress loop of the two tests: all green.
• Windows CI on this PR is the definitive check.

Follow-up

~30 other tests in this file carry the same latent cold-start Ask<SessionJoined>(3s) pattern. Scoping this PR to the two observed flakes; a base-class WarmUpSession(...) helper sweep is a good tracked follow-up.