nx_failclass.nx

nx_failclass.nx -- CLASSIFY THE ESTATE'S OWN FAILURE RECORD AGAINST A PUBLISHED TAXONOMY. The estate banks ~70,100 distilled law rows and every one of them is failure-shaped, but they were never CLASSIFIED and never COUNTED -- so nobody could say which failure class bites most, which is chronic, or whether any class is extinct. Anecdotes without a taxonomy. This organ answers that question over the WHOLE population, never a sample: nx_failclass scan <signatures.conf> <learnings.jsonl> <worklist.out> nx_failclass selftest AND, since 2026-08-20, THE FM11 RUNG OF /compare/failmodes -- a second census over a DIFFERENT population, the refusals this estate actually EMITS: nx_failclass refusal <refusal_shape.conf> <worklist.out> nx_failclass refusal_ratchet <worklist> <baseline> The ruler lives in nx_refusal_shape_lib so this verb and nx_refusal_shape_gate cannot disagree, and it scores the EMITTER only: whether the message names its subject and its remedy. Whether a caller HEEDED it is a second, independent axis, census-only over the call record, and no counter in this organ may be read as evidence about it. THREE PROPERTIES IT REFUSES TO GIVE UP: 1. u2605 IT IS A CENSUS, NOT A GATE. An uncalibrated classifier must report NUMBERS, never verdicts. There is deliberately no verdict= line and no RED. Exit is 0 for a completed scan, 3 for an unreadable input, 2 for usage. A reader who wants a bar must set it themselves against numbers they can see. 2. u2605 THE REASON TRAVELS WITH THE COUNT. Every counted row is also WRITTEN to the worklist with the exact signature that classified it. A count without a worklist is not actionable, and a worklist without the reason is still one step short -- the reason is already in hand at measure time and is normally thrown away. 3. u2605 UNCLASSIFIED IS ITS OWN BUCKET AND MULTI IS ITS OWN AXIS. A row matching no signature is never folded into a known class; a row matching two classes genuinely names two, so it is counted once in the MULTI partition member AND once per class on the overlapping per-class axis, which is labelled as an axis and never summed into the partition. The partition PRINTS ITS OWN SUM: single + multi + unclassified must equal rows. WORD BOUNDARIES ARE LOAD-BEARING, NOT COSMETIC. Measured 2026-08-20 over 70,100 rows, the bare substring "cap" hits 7,802 rows and the WORD cap/caps/capped hits 2,992 -- the residue is "capability", "capture", "capsearch", "capacity". A substring matcher would have published a