Greptile Summary

This PR bundles three unrelated fixes from separate issues: launchd restart throttling (#27590), usage token normalization (#27509), and chokidar v5 compatibility (#27404). According to CLAUDE.md: "Group related changes; avoid bundling unrelated refactors." Each fix addresses a different subsystem and should ideally be in separate PRs for easier review and rollback.

Changes included:

• Adds ThrottleInterval: 60 to launchd plist to prevent rapid gateway restarts when lock files are stale
• Fixes token counting for providers that return input_tokens: 0 with actual counts in prompt_tokens (yunwu-openai)
• Adds glob dependency (v13.0.6) for chokidar v5 compatibility

All three fixes are technically sound with appropriate test coverage. The launchd throttle interval correctly addresses the 10-second restart loop by waiting 60 seconds (exceeding the 30-second stale lock threshold). The usage normalization logic properly prefers non-zero token counts. The glob dependency addition has no visible code changes in this PR.

Confidence Score: 4/5

• Safe to merge - all code changes are correct with proper test coverage
• Score reflects that the code implementations are sound and well-tested, but the PR bundles three unrelated fixes that would be better as separate PRs. The launchd throttle logic correctly solves the restart loop, usage normalization handles provider quirks properly, and the glob dependency is a straightforward addition. No logical errors or security concerns identified.
• No files require special attention - all changes are straightforward

_{Last reviewed commit: 5090556}

Landed via /landpr flow.

• Gate: pnpm check && pnpm build && pnpm test -- src/daemon/launchd.test.ts
• Land commits: 431726dff143b5e7998f370204d4ce7cccf6f4a3, db3c5a52b29d6f7f3bd3a5cc5f0e190d658ac4bc
• Merge commit: 7f863e22b06f7fdcaece417299db7a8fff4b5214

Unified changelog entry included for the restart/orphan loop cluster (#27605, #27590, #26904).

Thanks @kevinWangSheng!