nx_xport_lib.nx

nx_xport_lib.nx -- DC12: ONE AUTHORED ASSET INTO FOUR CONSUMERS, MEASURED BY ONE RULER. WHY THIS EXISTS: the /compare/dcc board claims a single authored asset serves storefront, game, VR and clinical use. Today that is a PLAN. Nothing measures it, so a consumer could silently receive a decimated, re-scaled or re-wound mesh and no instrument in the estate would say so. This lib turns the claim into a number per consumer. THE RULER IS md_measure AND THERE IS NOT A SECOND ONE. nx_mmdev_lib already measures exact point-to-triangle deviation in tenths of a millimetre and is gate-proven to read a known 5.0 mm lift as 5.0 mm and a self-compare as zero. A new deviation metric on this board would be a duplicate ruler, i.e. a defect. Everything below COMPOSES it. The exporters are composed too -- write_glb and write_obj are called, never re-implemented -- so what is measured is the artifact the real storefront lane ships. WHAT IS MEASURED, PER CONSUMER: the artifact is written, then the geometry is RECOVERED BACK OUT OF THE ARTIFACT BYTES (not from an in-memory copy that never crossed the format) and measured against the source in BOTH directions. Deviation alone is not sufficient and this lib says so in code: a mesh mirrored about its own centre has ZERO point-to-triangle deviation and is still wrong, so orientation (the sign of the signed volume) and scale (per-axis AABB extent) are separate axes, not decoration. THE BOUND IS DERIVED, NEVER PICKED. Every consumer holds coordinates in its own integer unit. A round trip through a unit of U integer steps per metre, from the ruler's NM_UNIT_PER_M steps per metre, spends at most half a step in each direction, and NOTHING AT ALL when U is an integer multiple of the ruler's unit (the forward map is a multiplication and the inverse an exact division). The glTF path adds a second, computed term: write_glb encodes each integer coordinate with gl_i2f32, which TRUNCATES the mantissa, so the loss is the float32 ulp AT THIS ASSET'S OWN COORDINATE MAGNITUDE -- computed from the mesh, zero while the magnitude fits the 24-bit significand, and rising by itself if it ever does not. ABSTENTION IS NOT A SCORE OF ZERO. A consumer that cannot be measured reports UNREACHED with a named reason. A consumer that silently reads zero deviation is indistinguishable from one that was never measured, which is the whole failure this rung exists to prevent. DECLARED IMPRECISIONS, so no reader takes them as exact: (1) The volume-magnitude tolerance uses the AABB as a proxy for surface area. A convoluted mesh has more area per unit volume than its box, so on such a mesh that tolerance is OPTIMISTIC. It is a SUPPORTING axis only; the load-bearing orientation test is the volume SIGN, which needs no tolerance and cannot be optimistic. (2) The signed volume is the enclosed volume only for a CLOSED mesh. For an open mesh it is still a valid like-for-like shape statistic, because source and recovered are evaluated with the SAME origin (the source AABB minimum), but it is not a volume. (3) md_measure's own declared imprecisions carry through: its mean is valence-weighted and its p95 is