fix(desktop): reduce idle polling and checkpoint turns

[wqshan-cn]

Summary

Improve two desktop reliability paths:

• Closes [Bug]: 桌面端会话持久化缺陷——仅TurnDone异步保存，长对话中间状态与版本更新时数据丢失 #9719: checkpoint durable transcript state after the user message is actually appended, instead of using TurnStarted as an unsafe proxy, and make TurnDone a real persistence barrier.
• Refs [Bug]: 桌面端空闲时 CPU 持续占用约 1.2–1.5 个核心（reasonix-desktop 忙轮询，多个版本持续存在） #9539: reduce redundant BackgroundRuntimes bridge polling and prevent overlapping requests. The session-catalog repair-loop CPU root cause is handled separately by merged PR fix(sessioncatalog): stop idle CPU repair loops / 根治空闲历史修复满核 #9697; this PR does not claim to fix that loop itself.

Changes

• Keep TurnStarted as a lifecycle signal, then emit turn_phase=working only after the user message is in the session; desktop autosave checkpoints that post-append boundary.
• Keep streaming deltas out of autosave while checkpointing tool results, prompts, compaction completion, committed user-turn state, and TurnDone.
• TurnDone persistence barrier: Emit(TurnDone) now waits until (a) the autosave loop drains its current write plus any queued saveAgain pass (waitForTabSnapshot), and (b) the flushed display-only text finishes its current persist attempt including bounded retries (waitForPendingDisplayWrites). A process exit immediately after a completed turn can no longer lose the final transcript. Mid-turn checkpoint events stay asynchronous — no per-token disk writes.
• Add a single-flight background runtime refresh coordinator.
• Stop fallback polling when no background runtime is active; use a 5-second fallback while work remains.
• Retry transient bridge failures with bounded exponential backoff and handle internal Promise rejections.
• Order-insensitive runtime comparison: the backend builds BackgroundRuntimes from Go maps whose iteration order is unstable, so sameBackgroundRuntimeLists now compares by tabId instead of by array position; identical runtimes no longer trigger redundant React state updates.
• Prevent overlapping WorkspaceConflictForTab requests and lower its fallback cadence to 1.5 seconds.
• Ignore late results after coordinator disposal and avoid redundant React state updates.
• Add regression tests: session content is readable from disk immediately after Emit(TurnDone) returns (TestTurnDoneWaitBlocksUntilSnapshotWritten), and reordered identical runtime lists compare equal.

Verification

• go test ./internal/agent ./internal/control -count=1 passed (after rebase onto upstream/main-v2@c01ec347b).
• Focused desktop autosave tests passed, including the new persistence-barrier regression test, TestScheduleSnapshotCoalesces, TestCloseTabNoResurrectionFromAutosave, and the cannot-persist guard tests.
• background-runtime-refresh.test.ts passed (including the new order-insensitivity assertion).
• tsc --noEmit passed after regenerating the local (gitignored) Wails bindings for this branch — earlier failures came from stale generated bindings left behind by the feat(subagents): allow per-profile max-steps cap #9659 work, not from PR code.
• ESLint passed for changed frontend files. git diff --check passed.

Documentation-impact: none - this change affects internal desktop polling and persistence behavior; existing user-facing documentation does not describe these implementation details.

Cache-impact: none - no provider prompt, tool schema, or cache-prefix behavior changes.

Cache-guard: go test ./internal/agent ./internal/control -count=1, the focused desktop autosave tests (including the new TestTurnDoneWaitBlocksUntilSnapshotWritten barrier regression), and background-runtime-refresh.test.ts cover checkpoint ordering, the TurnDone persistence barrier, order-insensitive runtime comparison, refresh coalescing, retry behavior, and disposal.

System-prompt-review: no system-prompt or agent-instruction templates are touched; all changes are desktop event-sink persistence, autosave scheduling, and frontend refresh coordination.