nx_targetfit.nx

nx_targetfit.nx -- DID THE GENERATOR HIT ITS TARGET, AND BY HOW MUCH DID IT MISS. THE QUESTION THIS ANSWERS, AND WHY IT IS NOT ONE WE ALREADY ASK. The estate owns two rulers that look adjacent and answer different things. nx_mmdev_lib md_measure gives exact per-vertex deviation against a reference -- how far apart two SHAPES are. nx_proportion_guard pg_check gives absolute plausibility -- whether a body is a POSSIBLE body at all. Neither answers the operator's question, which is: we ingested a reference, we derived a target from it, the generator produced something -- DID IT HIT THE TARGET, while still being allowed to vary. THE WHOLE DIFFICULTY IS THE WORD 'ALLOWED'. A conformance ruler that demands exactness destroys the thing it is measuring: procedural generation exists to produce variety, and a checker that passes only near-identical output would drive the generator toward a single figure. Rule 25 is explicit -- build intelligence, never strip features; over-efficiency destroys variety and richness. So the verdict here is a BAND, never a point, and the band's width is published beside every verdict. THE BAND IS NORMALISED BY THE AXIS'S OWN PLAUSIBLE ENVELOPE, WHICH IS THE PART THAT MAKES IT MEAN ANYTHING. A naive ruler expresses tolerance as a percentage of the target: five percent. But five percent of total height is about 87 mm and five percent of breast projection is about 4 mm, and those are not comparable quantities -- one is a shrug, the other is a different body. So a miss here is expressed as PERMIL OF THE AXIS'S PLAUSIBLE RANGE, taken from pg_bound. A tenth of the human range means the same thing on every axis, which is the only way a multi-axis verdict can be summed without secretly weighting whichever axis happens to have the biggest numbers. WHAT IS DERIVED AND WHAT IS DECLARED, stated rather than blurred: the NORMALISATION is derived -- it comes from bounds nx_proportion_guard already holds and defends. The BAND EDGES (tight, slight, moderate) are DECLARED DEFAULTS, not measured, and they are parameters so a caller can drive them from conf. Anyone who tightens them by feel rather than by measurement is doing the thing that file's own provenance note forbids. AND THE ENVELOPE'S REFUSAL PROPAGATES. pg_bound REFUSES an axis with no bound row, on the principle that missing is not permission. An axis we cannot normalise is UNMEASURABLE here, never a pass: silently treating an unbounded axis as conforming would let the axes nobody has characterised be exactly the ones that always agree. license_tier: ORIGINAL No hardware writes (Rule 26). Pure integer, pure buffer.