[zxch3n]

Related issue

Problem / pressure

The first CLI attempt of run 40630618794 timed out in the disconnected tool_call_update case; rerunning the job passed. The fixture froze timeout timers while loro-repo 0.25.0 measured persistence work with real performance.now(). A persistence yield could be captured by the fake clock, then discarded when cleanup switched clocks, stranding repo.destroy().

Summary

Keep one real timer clock for these repository lifetimes. Use explicit prompt/disconnect signals, synchronously cancel the scheduled batch flush, and exercise both buffered and explicitly persisted notifications before disconnecting. Preserve all existing assertions and the 30-second timeout; no retry or product-code changes. Add a bilingual testing note and correct the fixture clock comment.

Visual explanation

Before: persistence exceeds 5ms → fake setTimeout(0) → uninstall fake clock
                                                       ↓
                                        callback lost → repo.destroy waits

After:  notification → cancel batch timer → prompt-entered signal
        → leave buffered / await explicit flush → release disconnect
        → await execution, drain and repo close on the same timer clock

Before / after

Before	After
Persistence cleanup depends on whether real work exceeds its budget while fake timers are installed.	SDK persistence yields remain runnable through cleanup.
One implicit notification/disconnect order.	Explicit gates cover buffered and persisted notifications for all three update types.

Test plan

• Entire file: 20 consecutive independent Vitest processes, 74 cases each (1,480 passed), with GIT_EXEC_PATH cleared for synthetic Git fixtures.

• Controlled slow branch: temporarily install a hoisted monotonic clock advancing 10ms per read before the SDK captures it. All six disconnect variants passed in five independent processes (30 cases); timer stacks confirmed DocManager.drainDirtyDocs yielded.

• Ablation: remove only the real-timer fix under that injected clock. Execution, drain and assertions complete, then cleanup reproduces the original 30-second timeout.

• The original CI log establishes a timeout, not its exact blocked await; the controlled reproduction establishes the fixture defect.

• pnpm format, changed-file Oxfmt check, pnpm run docs check, workspace typechecking, both lint passes, i18n and all repository boundary guards passed.

• pnpm check reached script tests, where the sandbox denied listen 127.0.0.1 (EPERM). Re-running pnpm test:ci with local-port access passed scripts and 6,405 component cases; CLI reported 5,006 passed, 4 skipped and one failure in the unchanged github-git-transport.test.ts recursive SSH fixture (context_unreadable). The changed batching file passed in that full CLI run. Direct Node execution with GIT_EXEC_PATH removed passed that Git fixture on both unmodified main 9fec3a21 and this head. Full pnpm check is therefore not claimed green; downstream package tests after the recursive-run failure did not finish.

Context handoff

Same-repository maintainer branch. Fixture-only change; no live-provider or packaged-desktop acceptance is claimed.