nx_swarm_discover.nx

nx_swarm_discover.nx -- SF-DISCOVER: find GPU inference capacity on the LAN and register it. WHY. The swarm's endpoint SSOT (knowledge/swarm_nodes.conf) is HAND-MAINTAINED, and that is how the estate lost its GPU for weeks: DHCP moved the laptop .193 -> .192 and five hardcoded copies went stale in the same instant while every swarm gate still read GREEN. Binding the five copies to one row fixed the DUPLICATION; it did not fix the fact that a human still has to notice and type. ★★★★★A REGISTRY THAT ONLY A HUMAN CAN UPDATE GOES STALE AT EXACTLY THE MOMENT IT MATTERS -- when hardware moves, which is the same moment nobody is watching. MEASURED 2026-08-04, and it is the reason this exists: a sweep of every live host on 192.168.8.0/24 found EXACTLY ONE inference endpoint (the laptop 5080 at .192:7861). The "west dual 3090" row in the SSOT pointed at 10.0.4.13 -- a different, unreachable subnet -- so the estate believed it had 48GB of VRAM it could not touch. ★A CONFIG ROW IS A CLAIM ABOUT THE WORLD, AND AN UNVERIFIED ONE IS JUST A WISH WITH AN IP ADDRESS. WHAT IT DOES: probe a host range for real inference endpoints, VERIFY each by asking it to identify itself, and report capacity. With `apply`, APPEND newly-found endpoints to the SSOT. ⛔IT NEVER DELETES OR REWRITES AN EXISTING ROW. A discovery pass that can silently drop a node is a worse failure than a stale row: the stale row fails loudly at dispatch, a deleted one vanishes. Absent hosts are REPORTED, never auto-removed -- a host being off is not proof it is gone. nx_swarm_discover scan [a.b.c] -- probe .1-.254, report what actually answers nx_swarm_discover apply [a.b.c] -- same, then APPEND unregistered finds to swarm_nodes.conf license_tier: ORIGINAL expect_exit: 0 module: nishi-core.swarm.discover