Shared data directory on filesystems with non-exclusive advisory locks: detection and guidance #45

The exactly-once logic for processes sharing a data directory (#32) relies on advisory file locks. On some mounts locks succeed but do not exclude across hosts/VMs (network filesystems mounted without locking, some container bind mounts from a desktop VM host, WSL2 drive mounts). The result is silent double delivery. Add detection where possible (e.g. a cross-handle probe at startup), a one-line warning, and a documented recommendation (one SDK host per directory there). Also: heartbeat gaps in the lock-less fallback — ensure_marker can restore the live marker without recreating its heartbeat file after the tree was wiped, and a single recovery step or before_send longer than the heartbeat staleness window does not refresh it.

🤖 Generated with Claude Code