Defer throttled streaming chunks via pending-flush timer

[eanzhao]

Summary

TurnStreamingReplySink previously dropped any delta that arrived inside the throttle window until the next delta showed up. When the LLM stalled near the window edge (cold-start, router handoff, tail-of-thought), the user saw the reply freeze for the rest of the throttle interval before the next batch landed.

This PR:

• Stashes the latest accumulated text on each in-window delta and arms a one-shot deferred flush timer that publishes the stash when the window closes. Multiple in-window deltas collapse onto the latest text.
• Reflushes after in-flight dispatch (openclaw FlushController-style): while a dispatch is in flight, additional deltas — caller-driven or timer-driven — are stashed; the dispatch loop reflushes the most recent _pendingText after the in-flight chunk completes. The conversation actor still observes strict edit ordering.
• Makes the sink IDisposable so ChannelLlmReplyInboxRuntime cleans up an unfired timer on paths that skip FinalizeAsync (interactive reply, empty reply, exception). Wired via a using declaration.
• Adopts Microsoft.Extensions.TimeProvider.Testing.FakeTimeProvider for deterministic timer-driven test control. The hand-rolled FakeTimeProvider in TurnStreamingReplySinkTests is replaced with the BCL-aligned one.

Background: studied https://github.com/ColinLu50/openclaw-lark-stream — src/card/flush-controller.ts (mutex + reflush-on-conflict) and streaming-card-controller.ts (deferred flush after long gap). This PR adopts the deferred-flush + reflush patterns; it does NOT adopt the OpenClaw CardKit 2.0 streaming card path (we stay on /reply + /reply/update edit-message via NyxID relay).

Out of scope

Followed up by #405:

• Replace Disabled + SuppressInterim boolean pair on NyxRelayStreamingState with an explicit phase enum + PHASE_TRANSITIONS table.
• Centralize message-unavailable / recall short-circuit on ConversationGAgent so future handlers don't have to duplicate the checks.

Test plan

• dotnet build sink test project — 0 errors.
• TurnStreamingReplySinkTests — 12/12 pass (7 retained, 5 new: deferred timer flush, finalize-cancels-timer, dedup-against-last-emitted, dispose-prevents-flush, dispose-idempotent).
• Full Aevatar.GAgents.ChannelRuntime.Tests — 166/166 pass.
• Aevatar.GAgents.Channel.Protocol.Tests — 119/119 pass.
• tools/ci/test_stability_guards.sh — passed (no new Task.Delay / WaitUntilAsync; uses TimeProvider.CreateTimer + FakeTimeProvider.Advance).
• CI: Channel slice + Distributed slice.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 94409bd382

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Wait for in-flight dispatch when finalizing stream

When _dispatchInProgress is true, this path only stashes _pendingText and returns immediately, so FinalizeAsync can complete before the final chunk is dispatched. ChannelLlmReplyInboxRuntime.ProcessAsync then sends LlmReplyReadyEvent right after awaiting FinalizeAsync, and if that ready event is handled first, later chunk events are treated as late and dropped in ConversationGAgent.HandleLlmReplyStreamChunkAsync (the processed-command guard). This makes final-stream delivery race-dependent and can skip the intended final edit path.

Useful? React with 👍 / 👎.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 70.38%. Comparing base (d102e47) to head (07084b9).
⚠️ Report is 5 commits behind head on dev.

@@            Coverage Diff             @@
##              dev     #407      +/-   ##
==========================================
- Coverage   70.39%   70.38%   -0.01%     
==========================================
  Files        1175     1175              
  Lines       84452    84452              
  Branches    11124    11124              
==========================================
- Hits        59446    59441       -5     
- Misses      20715    20720       +5     
  Partials     4291     4291              

Flag	Coverage Δ
ci	70.38% <ø> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.
see 3 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.