nx_cwdguard.nx

nx_cwdguard.nx -- THE STANDING GUARD for the CWD-DEPENDENT-VERDICT class (debt 1785879363). WHAT BIT US (measured 2026-08-04, three separate organs in one session): * nx_swarm_endpoint_lib read knowledge/swarm_nodes.conf relative; the gen daemon runs with CWD=/volume1/ai/gen, so the SSOT silently missed and gen dispatched at a DEAD GPU address. * nx_writebench printed verdict=GREEN coverage=553 from buildroot and result=LIAR claimed-artifacts-missing=8 from nishihost -- the tools-daemon CWD, i.e. EVERY MCP call. * nx_law_warden returned RED "law plane unseeded" and nx_sota_status reported an EMPTY estate (no proven domains at all) purely because of the directory they were launched from. ★★★★★★A VERDICT THAT CHANGES WITH THE CALLER'S WORKING DIRECTORY IS NOT A MEASUREMENT. ★WHY THIS DETECTOR IS STATIC, NOT A DIFFER. The obvious guard -- run each tool from two directories and compare -- is WRONG HERE, and I proved it on myself: comparing output md5 flagged nx_barcheck (a TLS bench whose ms= timings vary per run) and nx_frontier_watch (which APPENDS a row per run, so its base row advances every time) as CWD-fragile when NEITHER IS. It also would EXECUTE mutating organs twice per sweep just to look at them. So this guard reads SOURCE: an organ is AT RISK when it opens an estate path RELATIVE and carries no anchor. That is cheap, deterministic, side-effect-free, and has no false positives from clocks or counters. ⚠HONEST ENVELOPE: at-risk is a RISK, not a proven defect -- an organ whose subject genuinely IS "whatever tree I was pointed at" is CORRECT to read relative. This names candidates for a human or a follow-up differ to adjudicate; it must never be read as a list of confirmed bugs. nx_cwdguard -- scan, print the at-risk roster + verdict nx_cwdguard selftest -- teeth, incl. the NEGATIVE control (an anchored organ is NOT flagged) license_tier: ORIGINAL expect_exit: 0 module: nishi-core.estate.cwdguard