🧹 fix: Close Native Scratch Directory Streams

[danny-avila]

Summary

I made native scratch recovery close every directory stream deterministically and updated the remote bridge lifecycle documentation for negotiated worker concurrency.

• Replace async directory iteration with explicit bounded reads and an unconditional close, including traversal-limit and inspection failures.
• Preserve descriptor-relative recovery, traversal limits, and existing symlink-race protections.
• Document bounded worker lease slots, same-isolation-key serialization, and concurrent conversation worktrees.

Change Type

• Bug fix (non-breaking change which fixes an issue)
• Documentation update

Testing

• npm run build in packages/code
• node --trace-warnings --test dist/native-sandbox.test.js (61 passing; no FileHandle garbage-collection warnings)
• Local three-process canary using current LibreChat dev, local Code API, and a native worker with two lease slots and conversation worktrees:
  • Derived two independent workspace instance IDs from separate LibreChat conversation IDs.
  • Ran two four-second commands concurrently in 5.16 seconds total.
  • Verified distinct generated worktree paths and isolated persisted contents.
  • Verified the source repository remained unchanged.

Test Configuration:

• macOS
• Node.js 24.16.0
• Redis 8.4.0
• Code API local mode with two negotiated workspace slots
• Native SRT worker with conversation-scoped Git worktrees

Checklist

• My code adheres to this project's style guidelines
• I have performed a self-review of my own code
• I have commented in any complex areas of my code
• I have made pertinent documentation changes
• My changes do not introduce new warnings
• Existing focused tests demonstrate that my changes are effective
• Local unit tests pass with my changes