PID-based file semaphores in ~/.claude/nf-sem/ leave stale locks on crash #176

Context

bin/provider-concurrency.cjs implements per-provider concurrency limits via PID-files in ~/.claude/nf-sem/. The cleanup is best-effort:

• If a dispatched process exits cleanly, it releases its slot file
• If the process crashes, gets killed (SIGKILL), or the parent shell terminates abruptly, the lock file persists with a stale PID
• Subsequent slot acquisitions can be incorrectly blocked until a manual cleanup or until the file's stale-PID check fires

Surfaces as "all slots UNAVAIL with no clear reason" during stress / interrupted runs.

Flagged in .planning/quorum/debates/2026-05-13-is-the-quorum-robust.md by opencode-1 as bounded weakness #4.

Proposed solution

Switch to fs.flock-style advisory locking (POSIX/Windows-bridged via proper-lockfile or similar). OS-level locks release automatically on process exit regardless of how the process dies.

If POSIX-only is acceptable: use fcntl.flock via the @petamoriken/float16-style native binding, or flock(1) as a subprocess wrapper.

Fallback: extend the existing stale-PID check to run on every acquisition attempt, not just on file creation. A PID that doesn't exist in process.kill(pid, 0) → release the lock and continue.

Acceptance criteria

• Killed processes (SIGKILL, parent shell exit) don't leave stale locks
• No manual rm -rf ~/.claude/nf-sem/ needed for recovery
• Concurrency limits still enforced under heavy load
• Cross-platform (macOS + Linux at minimum)