test: close statically proven CI race causes

[SunSi12138]

Refs #387.

Scope

This PR groups only root causes that can be closed from the current source/CI evidence. Each root cause is kept in a separate commit; no timeout widening, blanket retries, quarantine, or production behavior changes are used as a substitute for causality.

1. Tracked unary deadline at emission

SharpLinkClientTrackedEmissionDeadlineTests.TimedUnaryDroppedAtEmissionShouldCompleteWithoutDeadlineTimerCallback advanced manual time immediately after starting InvokeUnaryAsync, then used a later send-queue flush as if it proved the target Request had reached emission. There is no ordering edge between invocation scheduling and that flush marker, so the marker can complete first; because the test intentionally does not run deadline timers, the invocation can then remain pending until the test watchdog.

The test now uses the existing TestTransportConnection.RunOnNextOutputBufferRequest seam. Before arming the one-shot hook, it drains output already associated with ConnectAsync through the ready session's semantic FlushSendQueueAsync barrier, matching the established #392 pattern and excluding the initial heartbeat Ping from consuming the hook. Manual time therefore advances only when the send pump has actually reached the target Unary Request's output-buffer acquisition, immediately before the production emission path re-samples the deadline. This preserves the assertion that the expired Request never reaches transport without adding a wall-clock timeout or retry.

2. Protocol phase gate flush scheduler bound

NegotiatedSessionOptionsTests.SendPumpShouldEnforceProtocolPhaseBeforeQueueing had already completed the phase-rejection checks and queued the allowed draining Response, then wrapped the semantic FlushSendQueueAsync barrier in an unrelated two-second wall-clock WaitAsync.

The flush itself is retained because the final queue-accounting assertion requires the allowed cleanup frame to drain. Only the local two-second scheduler clock is removed.

Static-audit rule

The remaining #387 records were rechecked against current dev; a production-path fix is included only when state/ownership/protocol behavior can be proven wrong from the source. Historical items already closed by merged PRs are not duplicated, and unresolved one-off evidence is not changed speculatively.

Review follow-up

• 2be9488d: establish a deterministic pre-existing-output barrier before the tracked-unary emission hook, so the first heartbeat Ping cannot consume the callback.
• Inline review thread resolved after applying the fix.

Validation

CI is running on exact head 2be9488d386fa31d01ee3306dabd77e13205e78c.

[SunSi12138]

整体两个 root-cause 方向都合理，第二个 phase-gate flush scheduler-bound 修复也有 #387 的 exact-head recurrence 证据。当前唯一 blocker 是 tracked-unary 新 hook 仍可能被 ConnectAsync 后立即启动的首个 heartbeat Ping 抢走，因此目标 Request 的 emission ordering edge 还没有完全建立，见 inline。

[SunSi12138]

上次唯一 blocker 已修复：tracked-unary 测试现在先通过 ready session 的 FlushSendQueueAsync 排空 ConnectAsync/首个 heartbeat Ping 的既有输出，再 arm one-shot output-buffer hook，因此 callback 才真正唯一绑定到目标 Unary 的 emission boundary。第二处 phase-gate flush 改动仍只移除语义 barrier 外层的 scheduler-bound timeout。exact-head Fast / Extended / CodeQL 均通过，未发现新的工程级问题。