fix(desktop): keep live cache averages in sync / 修复实时缓存平均命中率同步

[SivanCola]

Summary

• Refresh the shared all-source ContextInfo after each provider usage event during an active turn, so the desktop status bar no longer waits for turn_done to update its session-average cache hit rate.
• Make the status bar and context panel prefer the same live ContextInfo for their session-average cache metric, while keeping independently refreshed panel telemetry as a fallback.
• Reject stale async snapshots by checking the active tab, usage sequence, and turn lifecycle before updating shared state.
• Add integration coverage for executor and subagent usage, stale private panel snapshots, rendered status-bar and context-panel values, and out-of-order snapshot completion; register the delivery-mode stream tests in the standard frontend suites.

Verification

• pnpm test:all
• pnpm typecheck
• pnpm test:typecheck
• pnpm check:css
• ./scripts/cache-guard.sh
• git diff --check

Cache impact

Cache-impact: none — this only refreshes existing local telemetry; provider-visible prompts, tools, and request serialization are unchanged.
Cache-guard: ./scripts/cache-guard.sh passed all scenarios with tail averages of 92–95% against the 90% threshold.
System-prompt-review: N/A