nx_daemon_supervisor.nx

nx_daemon_supervisor.nx -- SOVEREIGN data-driven daemon supervisor. THE organ that makes daemon lifecycle the ecosystem's job, not Claude's: a daemon is a REGISTRY ROW (data), not compiled into a guard's target list and not nursed by a bash loop. Reads daemons.reg HOT every cycle, responds-health probes each via hp_probe (SERVING / HUNG / REFUSED -- catches the wedge that PID/port checks miss), and REVIVES the dead by fork+execve of the declared argv. 2-consecutive-fail discipline (in-memory, keyed by name) so a single transient probe never triggers a relaunch. On HUNG (wedged, holding the port) it kills the matching process first (the new instance could not bind otherwise), then launches. Zombies reaped every cycle (wait4 WNOHANG). Add a daemon = a registry line + reload -- NO recompile, NO shell. BOUNDED-VSZ DISCIPLINE (2026-07-16): every reusable buffer is allocated ONCE before the loop and reused each cycle. The hot loop performs ZERO sys_mmap. The prior version mmap'd per cycle -- fatally, it read the registry with sys_read_file, whose 4 GiB default cap (used whenever lseek can't size the file) maps 4 GiB of VSZ per cycle and never frees it. That drove VSZ to 116 GB in ~29 cycles, and under heuristic overcommit fork() then returns -ENOMEM (proven: REVIVED ... pid=-12), so revives silently stopped. The registry is now read into a fixed buffer via sys_read; ds_decw is temp-free; VSZ stays flat forever. daemons.reg (pipe-delimited; lines starting # or blank ignored; args are space-separated, no spaces-in-args): <name> | <cwd> | <argv0 [arg1 arg2 ...]> | <health_port> | <arm: revive|watch> health = hp_probe(port): ANY HTTP response on 127.0.0.1:port = healthy (401/404 count). revive = relaunch (and kill-the-wedge) on 2 consecutive non-SERVING probes; watch = detect+log only. license_tier: ORIGINAL module: nishi-core.ops.daemon_supervisor