nx_raidwatch.nx

nx_raidwatch.nx -- ARRAY + KERNEL-IO HEALTH BEAT. The estate had NONE. WHY THIS EXISTS (2026-08-07). A grep across every organ for mdstat, smartctl or raid returned only substring false positives (a bip39 wordlist, a chemistry table). Nothing in a 140 TB estate was watching whether its arrays were intact. Two live faults were sitting there unwatched: md3 (raid1, the NVMe mirror) DEGRADED [2/1] [_U] -- one member gone, /dev/nvme0n1p1 has no device node at all, so it cannot even be re-added without physical intervention. It was recorded in a memory file and NOTHING CHECKED IT, which is the difference between a note and a monitor. sata8 throwing repeated UNC (uncorrectable) read errors on ONE sector, 840 kernel records over four hours, while the array it backs still reports [8/8] because raid5 keeps reconstructing the block from parity. A disk can be failing loudly while every array metric reads healthy. ROOT-FREE BY CONSTRUCTION, because the beat runs as the estate user: /sys/block/<md>/md/degraded, raid_disks, level, mismatch_cnt -- structured integers, no text parsing of /proc/mdstat, so this cannot be broken by a format change. /dev/kmsg -- readable here, opened O_NONBLOCK so the final read returns EAGAIN instead of BLOCKING FOREVER at the end of the buffer. A monitor that can hang is not a monitor. smartctl is present but needs root (open of /dev/sata8 = Permission denied, measured), so per-disk SMART is deliberately OUT OF SCOPE here rather than silently half-working. REFUSES if it can read neither surface: ignorance is not health, and a GREEN produced by an unreadable sysfs would be worse than no beat at all.