nx_netprobe_ladder.nx

nx_netprobe_ladder.nx -- THE DISCRIMINATING EXPERIMENT for the shared TCP liveness probe. THE SITUATION, measured 2026-09-05: nx_gen's dispatch selector and nx_swarm_heal BOTH report the laptop image worker DOWN, while nx_worker_dispatch gets HTTP 200 from GET /v1/models at the SAME address. That looks like two witnesses against one -- it is not. nx_netprobe_lib's own header states both callers were consolidated onto np_probe_up, so the two DOWN verdicts are ONE WITNESS COUNTED TWICE, and they share whatever blind spot it has. WHAT THIS VARIES AND WHAT IT HOLDS FIXED: the address is FIXED, only timeout_ms varies. A flip from DOWN to UP as the bound grows names the timeout as the cause; NO flip refutes the timeout hypothesis and points at the probe's mechanism instead. Running the experiment that can REFUTE the theory, before any fix is written, is the whole point -- a plausible cause confirmed by nothing is how this estate has burned entire sessions. The .193 line is a CONTROL, not a target: nx_swarm_heal calls that endpoint FRESH, so if the ladder reports it DOWN at every bound then the probe is broken for every host and the subject is exonerated. license_tier: ORIGINAL No hw writes (Rule 26).