nx_sdfrender_gate.nx

nx_sdfrender_gate.nx -- gate for the SMOOTH-UNION SDF body (sovereign software GPU, continuous surface). T1 the body renders (solid skin block) T2 per-pixel shading gradient (SDF-gradient normals -> real lighting) T3 full figure top->bottom T4 CONTINUOUS: the centre column is ONE fused run (the parts FUSED, not disjoint blobs) -- the whole point of smooth-union over the rasterized ellipsoids T5 THE CAPTURE ACTUALLY LANDS -- four teeth where there used to be one, see the block below T6 pre-integrated SSS (sr_sss_preint), the non-gameable quad =================================================================================================== WHY T5 IS FOUR TEETH NOW (debt 1787678931, MEASURED 2026-08-25 -- this gate WAS the defect). The old T5 was one line: write_png(fb, ww(), hh(), "knowledge/nx_sdfbody.png"); if diff == 0 { PASS "determinism + PNG" } It DISCARDED write_png's return and gated on DETERMINISM ONLY. Determinism is trivially satisfied by WRITING NOTHING TWICE, so the tooth could not distinguish a landed capture from no capture at all. PROOF IT DID NOT: two runs whose framebuffers provably DIFFERED (skin 37624 vs 37626, bands 10352/18040/9232 vs 10353/18042/9231) left knowledge/nx_sdfbody.png byte-identical at 827bc7718e15f0437a1600326e7bc7ec46a34dfc7c189d2a51607f7ccdf185a5 / 142629 B across BOTH -- while this gate printed "T5 PASS determinism + PNG". A whole rung shipped with no trustworthy before/after image. A TOOTH THAT DISCARDS ITS ACTUATOR'S RETURN CANNOT SEE THE ACTUATOR FAIL. DETERMINISM IS NECESSARY AND IS NOT SUFFICIENT, AND THE TOOTH THAT NAMES ONLY THE NECESSARY HALF READS EXACTLY LIKE A PROOF OF THE WHOLE. The four that replace it, each named for what it can refute: T5a the publish RETURNS a byte count and the ARTIFACT ON DISK IS THAT SIZE -- receipt AND artifact T5b determinism, RENAMED so it can never again be read as evidence the file was written T5c A/B: two captures whose SHADER PROVABLY DIFFERS must be DIFFERENT FILES. This is the tooth a no-op writer, a stale-file writer and an open-that-silently-failed all die at, and the one the old gate had no analogue of. T5d fixture-reached-the-condition: the two swatches' shader values really do differ, asserted ARITHMETICALLY and FIRST, so a T5c failure can never be blamed on a fixture that never varied. plus two neg-controls that must FAIL for the teeth above to mean anything. AND THE OWED ARTIFACT. T5c's A/B pair is not a synthetic fixture -- it IS the before/after this rung owes. nx_sdfrender's own shading site says "curv=0 reproduces the former fixed 120/60/26 widths bit-for-bit", so evaluating the SHIPPED sr_sss_preint over the whole (N.L x curvature) domain with curvature PASSED THROUGH gives the rung, and with curvature FORCED TO ZERO gives the pre-rung shader exactly. Two images, one shipped function, no second implementation to be wrong: knowledge/nx_sdfsss_before.png pre-rung fixed wrap (every row identical: curvature ignored) knowledge/nx_sdfsss_after.png the shipped rung (rows widen with curvature)