nx_gate_fixture_ratchet_gate.nx

nx_gate_fixture_ratchet_gate.nx -- A GATE MUST NOT SHARE ITS FIXTURE WITH A PRODUCTION BEAT. 2026-08-07, ws=segstore. SIBLING OF nx_gate_dry_ratchet_gate, deliberately built to the same shape: same corpus, same baseline-file convention, same first-run honesty, same gv_bite non-vacuity cell. Only the TOOTH differs. Matching the known good was the point -- a second ratchet that behaved differently would be a second thing to remember, and the whole value of a ratchet is that nobody has to remember it. WHAT IT RATCHETS. A gate that keeps its scratch planes under knowledge/store/ is sharing a directory with the nx_segguard beat, which sweeps every 600s and folds any plane at or above its threshold. MEASURED 2026-08-07T10:01:56: segguard folded `segracegate-real 23 -> 1` WHILE nx_segrace_gate was mid-run, and the gate then reported "COMPACTION STILL ERASES COMMITTED SEGMENTS -- lost 1". A controlled comparison settled it: the SAME code on a /tmp fixture went GREEN 24/24, the pre-change baseline on the knowledge/store fixture went RED. The RED tracked the FIXTURE, not the code. ★★★★★★A GATE THAT SHARES ITS FIXTURE WITH A PRODUCTION BEAT IS MEASURING THE BEAT. ★A RED THAT A THIRD PARTY CAN CAUSE IS NOT A MEASUREMENT OF THE CODE UNDER TEST. WHY A RATCHET AND NOT A LINT-AND-FIX PASS: fixing the 9 gates that carry this today is O(LEAVES) and does nothing about the tenth someone writes tomorrow. The sibling gate says it best -- "a rule nothing must remember beats a list". This fails if they forget. SCOPE -- CORRECTED 2026-08-07, AND THE CORRECTION IS THE POINT. This header first claimed the defect was LATENT: "segguard has folded ZERO of them". That was measured over NINE planes I happened to have to hand, and all nine came back clean. Re-measured over the FULL population of thirty, the sweeper has touched TEN of them -- 57 logged COMPACT/FAIL/DEFER events, `ledgergate` alone 17 and `citatorgate` 9. So the defect is ACTIVE for a third of the population, not latent. ★★★★★★A SAMPLE THAT RETURNS ZERO IS THE EASIEST KIND TO BELIEVE AND THE MOST LIKELY TO BE UNREPRESENTATIVE -- 0/9 read as "no problem", the true rate was 33%. ★★★★★A LATENT DEFECT AND AN ACTIVE ONE ARE DIFFERENT FACTS -- AND YOU DO NOT KNOW WHICH YOU HAVE UNTIL YOU HAVE ASKED THE WHOLE POPULATION. THE TOOTH: count *_gate.nx sources that name a TEST-FIXTURE plane under knowledge/store/. current > baseline -> RED. Someone added a gate whose fixture the sweeper can rewrite. current < baseline -> GREEN, baseline TIGHTENS. The ratchet only ever closes. current == baseline -> GREEN, no change. DETECTOR, and why it is narrow ON PURPOSE: it fires only when the plane NAME following knowledge/store/ contains gate|test|probe|scratch|fixture. A gate that READS a real production plane is legitimate and must not be flagged; a raw "does the source mention knowledge/store/" check counts 124 gate files, of which only 9 are fixtures. The narrow form is the difference