nx_gen_dispatch_gate.nx

nx_gen_dispatch_gate.nx -- the REFEREE for gen G17 (dispatch honesty) on the CHAT path: a planted SLOW worker (listens, never answers) must make the LIVE daemon answer BUSY-TIMEOUT, and an ABSENT worker must make it answer the health-probe refusal -- and NEITHER may read "all in the registry are down", the false attribution measured 2026-08-29 (a socket watch showed the worker ESTABLISHED and rendering while the caller was told no worker existed). SUBJECT = the LIVE binary at /volume1/ai/gen (what serves, never a build). The daemon is launched on a throwaway port from a scratch cwd, so no swarm SSOT conf is visible and the fixture argv worker is its ONLY candidate; budget=2 (one probe connect, one POST) so it exits by itself and leaves no process behind. The slow fixture LISTENS AND NEVER ACCEPTS: the kernel completes the handshake into the backlog, so the daemon's probe and connect succeed and its read times out -- exactly the connected-and- silent worker the defect misattributed. Against the PRE-FIX binary this gate is RED by construction (the bite is the live history, not a planted mutant). RUNTIME ~4 min (three chat tries at the conf budget): run via nx_job_run, never the FAST gate_run path. G18 ADDED 2026-08-30 (the async lane): a third fixture ANSWERS, but only after DG_ANSWER_DELAY_MS. POST /api/chat_async must return a job id BEFORE that delay can have elapsed (so nothing waited on the worker socket), a poll must read pending first, and the planted reply must then land through GET /api/batch/<job> -- a completion slower than the submit round-trip lands instead of being dropped, which is the whole of G18's done-rule. The nonce in the landed reply is the isolation witness. license_tier: ORIGINAL expect_exit: 0