nx_staledir_gate.nx

nx_staledir_gate.nx -- proves the DIRECTION GUARD in nx_stale_check (F1145 / debt seq1008). WHY THIS GATE EXISTS: nx_stale_check's byte-compare tells you source and deployment DISAGREE but is structurally blind to WHICH IS AHEAD. The old code called every disagreement STALE, whose remedy text is "rebuild+restage" -- and for a binary that is AHEAD of its source that remedy DELETES the shipped feature. That is exactly how the nx_page_verify auto-connect-override would have been destroyed. So the guard is not cosmetic: it decides whether a human is told to run a DESTRUCTIVE command. The classifier under test is pure size-direction arithmetic, so it is testable without building anything: sd_verdict(fresh_bytes, deployed_bytes, identical) -> the verdict id. T1 identical -> CURRENT T2 fresh << deployed and tiny -> AMBIGUOUS (the pre-existing stub trap must SURVIVE this change) T3 fresh < deployed -> DEPLOYED-AHEAD <== THE NEW TOOTH: the seq1008 shape, remedy WITHHELD T4 fresh > deployed -> STALE (source genuinely ahead; rebuild IS the right remedy) T5 same size, differing bytes -> STALE (direction unknowable by size; must not claim AHEAD) T6 REGRESSION: the exact seq1008 magnitudes (572834 vs 574963) must classify DEPLOYED-AHEAD, NOT the stub AMBIGUOUS and NOT STALE -- a 2KB delta on a 575KB binary is the real case that slipped through the old stub-only guard. T7 ANTI-VACUITY: the guard must NOT swallow everything -- a genuinely stale target still reaches STALE, so the useful remedy is not lost. A guard that flags every case is as useless as none. expect_exit: 0 license_tier: ORIGINAL No hw writes (Rule 26).