Replace inotifywait with polling in cgroup cleanup

[assisted-by-ai]

Summary

Refactored the cgroup cleanup logic to use polling instead of inotifywait, eliminating dependency on the kernel's limited inotify watch pool and preventing ENOSPC errors on watch-constrained systems.

Key Changes

• Replaced inotifywait-based event monitoring with a simple polling loop that checks cgroup.events up to 50 times
• Removed file descriptor management and subshell complexity associated with inotifywait
• Simplified variable declarations by removing inotifywait_subshell_pid and inotifywait_fd, replacing with cgroup_drained and poll_count
• Added timeout detection with warning log when cgroup fails to drain within the polling limit
• Moved cgroup.kill write before the polling loop (more logical flow)
• Maintained the same 0.1 second sleep interval between polls for consistent behavior

Implementation Details

• The polling approach is viable because cgroup.kill sends SIGKILL, so draining is near-instant
• Maximum wait time is ~5 seconds (50 polls × 0.1 second sleep)
• Cleaner error handling with explicit cgroup_drained flag instead of relying on subshell exit status
• Reduced kernel resource consumption by eliminating inotify watch usage

https://claude.ai/code/session_01V4XkxCqroKpfDZE2Jknmdx

[ArrayBolt3]

Accepted in ArrayBolt3@36e9838.

[ArrayBolt3]

Accepted the followup commit in ArrayBolt3@ac45d4f.