- Add ToolUsage type import and use proper types in ClineProvider event handlers
- Make emitFinalTokenUsageUpdate() public in Task.ts for reuse
- Use emitFinalTokenUsageUpdate() in AttemptCompletionTool for consistent behavior
- Add final token usage emission in handlePartial for completion with commands

- Add hasToolUsageChanged() helper function to getApiMetrics.ts
- Add toolUsageSnapshot property to Task.ts to track tool usage changes
- Update saveClineMessages() to emit when either token OR tool usage changes
- Update emitFinalTokenUsageUpdate() to also check tool usage changes
- Add comprehensive tests for tool usage change detection

This ensures final tool usage stats are captured on task abort even if
token usage hasn't changed (e.g., when task is aborted before API request
completes but tools were already executed).

…tedAt in toolColumns deps

- Add emitFinalTokenUsageUpdate mock to clineC and clineB in nested-delegation-resume test
- Add usageUpdatedAt to toolColumns useMemo dependency array for proper recomputation when streaming tool usage updates

When tasks timeout, there's a race condition where the DB might not have
the latest stats when the frontend refetches. This fix adds fallback logic
to use streaming values (from the in-memory Maps) when DB values are
empty or missing.

Changes:
- taskMetrics useMemo: prefer DB values but fall back to streaming if empty
- toolColumns useMemo: prefer DB values but fall back to streaming if empty
- stats useMemo: same fallback logic for aggregate tool usage
- tool cells in table: prefer DB values but fall back to streaming if missing

After a task is aborted due to timeout, the extension rehydrates the task
with a new instance. This new instance has empty toolUsage, and if it emits
any TaskTokenUsageUpdated events, they would overwrite the final metrics
that were saved before the abort.

This fix adds a check to ignore TaskTokenUsageUpdated events once
taskAbortedAt is set, preserving the final metrics from before the abort.

Instead of ignoring TaskTokenUsageUpdated events after TaskAborted,
accumulate tool usage data using a MAX strategy. This ensures:
- Empty rehydrated data won't overwrite existing: max(5, 0) = 5
- Legitimate restart with additional work is captured: max(5, 8) = 8

This approach is more robust than simply ignoring post-abort events,
as it handles both spurious rehydration and legitimate restart scenarios.