nx_texpaint_gate.nx

nx_texpaint_gate.nx -- the gate for nx_texpaint_lib (/compare/dcc DC7, texture painting). THE HAZARD THIS GATE IS SHAPED AGAINST is the one the task names: a paint function that writes THE WHOLE ATLAS trivially passes "the known texel changed". So the load-bearing cell here is not the positive control on its own -- it is the pair of BOUNDED-WRITE teeth beside it, which count the changed texels against a HAND-COMPUTED disc area and then count the changed texels OUTSIDE the brush and require zero. A whole-atlas painter fails both, a do-nothing painter fails the first, and only a correct one passes both. Every disc area, every texel address and every painted BYTE VALUE below was worked out by hand before the code was run; they are written as literals with their arithmetic in the comment, so a fixture that agrees with a wrong implementation would have to be wrong in exactly the same way. THE ANTI-TEAR PROOF IS A PERMANENT NEGATIVE CONTROL, NOT AN ANECDOTE. tp_stroke_core is the naive painter -- shipped, public, callable -- so the tooth that proves the seam bleed matters RUNS the bleed-free version and shows the far side of the cut untouched, then runs the full stroke and shows it painted. That is a gv_bite that can be re-run by anyone, rather than a claim that the author once deleted the fix and watched a test go red. FIXTURE GEOMETRY, chosen so every boundary is exact and nothing rests on a rounding accident: res 64, grid 4 -> tile width in UV units 65536/4 = 16384, chart bounds 16384*64/65536 = 16 texels. So chart 0 is texels [0,16) x [0,16) and the four charts tile the atlas exactly, with no remainder. A SECOND map set at grid 3 is carried on purpose: 65536/3 truncates to 21845, the charts land at 0,21,42 and stop at 63, so ONE texel column belongs to no chart at all. That is the case where a partition claim could quietly leak, and it is asserted to SUM rather than assumed to. 100% sovereign. No hardware writes (Rule 26). license_tier: ORIGINAL expect_exit: 0