nx_registry_lock.nx

nx_registry_lock.nx -- the SHARED race-free lock PRIMITIVE, extracted (DRY, rule 15) from nx_registry_txn so every queue WRITER (reconcile / mark_migrate / express_lane + side-organs + the pulse/autorun daemons) can acquire ONE serializable, crash-safe, drvfs-portable lock around its read-modify-write of assignment_queue.tsv / row_markers.tsv. This is the hardware rung the operator asked for ("build the coordination from the hardware rung up so we can keep moving in parallel through daemons / orchestration"). PRIMITIVE: an O_CREAT|O_EXCL lockfile (the kernel guarantees exactly ONE creator wins -- no TOCTOU) stamped with the holder's acquire-epoch; a contender spin-serializes and STEALS a lock older than RT_STALE_SEC (a crashed holder, so the beat can never wedge). It deliberately does NOT use flock(2): flock is dead on /mnt/c drvfs AND the rv64->x86_64 syscall map (32->73) only activates on an nx_cc self-host redeploy (currently gated behind a RED cc_equiv run). O_EXCL needs neither -- it works today on drvfs with NO compiler redeploy. S-CLASS EXCEED over ad_acquire_lock (hard-fails on contention, no staleness recovery) and over atomic-rename-alone (saves a torn file but LOSES updates): this serializes contenders AND self-heals a crashed holder = serializable + crash-safe + drvfs-portable. MAINLESS LIBRARY: a consumer imports THIS file only (it pulls in nx_syscalls transitively) and must NOT also `import "nx_syscalls.nx"` directly -- the proven nx_assign_core<-nx_dep_audit convention, which dodges the double-import rc=6. Gate-proven (no separate gate authored) by nx_registry_txn's existing RTXN-GATE: fork 2 writers x N guarded increments -> lossless (==2N) while the unguarded control loses (<2N). license_tier: ORIGINAL module: nishi-core.autonomy.registry_lock depends: nishi-core.sys.syscalls capability: SHARED_REGISTRY_WRITE_LOCK_OEXCL