Independent POSITIONT numerical source acceptance, 2026-10-23

Historical source last updated, 2026-10-23. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.

ACCEPT scoped source/software behavior of complete candidate 5427a9e2e7d20984249cadcf67f1cb33cc9d85c0, relative to 931aff173131780f76d5116c4b53a181e2be2eab and main bb9c496ed13095cc832fdf607f0937fb9c14b7de. This is not merge or runtime qualification. Strict dedicated-key remote lookup independently matched all three revisions. No production changes in this review.

Why this admission is sufficient for the software path

Viewport integer origins [0,2048] and dimensions [1,2048] are exactly representable in f32 and agree with descriptor conversion. These are profile bounds, not complete draw validation: the software consumer additionally requires the viewport fit its target (maximum2048). Coordinates are finite because of their finite closed viewport bounds; z is in [0,1]. RHW must be a positive power of two in [2^-16,2^16]. For f32 inputs, Math.log2 cannot round an adjacent non-power-of-two float to an integer here; adjacent-boundary rejects also execute. Positive normal reciprocal W and recovered RHW are exact powers of two. Components can underflow, so reciprocal bounds alone are not the proof.

Native d3d_fixed_compile_stages (09aj, POSITIONT block) emits f32 viewport scale/bias DEF, RCP, MAD, MUL. The VM implements RCP by f32x4 division and MAD as separate f32x4 MUL then ADD. JS uses matching Math.fround boundaries, including bias products. Native projector (09ah d3d_software_project) divides clip coordinates by W, applies viewport operations with the same individual f32 roundings, and recovers reciprocal W. JS compares all recovered x/y/z/RHW to the input. Identity depth means native multiply-by-one/add-zero does not change ordered depth values; signed zero does not change coverage/depth tests. Finite clip-component and all six plane inequalities are checked independently.

Exact vertex round trips alone would be insufficient. Here all vertices are also inside every homogeneous plane with positive W, and user clipping is excluded. The production clipper consequently copies original vertices without constructing intersections or deleting homogeneous-zero vertices. It retains triangle order/edge membership. Projected screen position, depth and reciprocal W therefore reach the same rasterizer as the old screen-space path. Fixed color/UV MOV outputs and their subsequent multiply by reciprocal W are unchanged; interpolation uses the same f32 barycentric weights, interpolated reciprocal, reciprocal multiplication, texture sampling, quantized depth and coverage rules. There is no new JS interpolation implementation. Independent full-frame and analytic controls below exercise this reasoning beyond the old constant-red samples, including unequal extreme RHW, sloping depth and fractional edges.

Draw ranges, ownership and fallback

Reused accepted unrelated source evidence from reviewd5ba96fc and its prior review6ab1cff6, after verifying the unchanged materialized production blobs. The host already applies StartVertex/BaseVertex/min-index offsets and copies declared stream ranges into the exact packed snapshot; INDEX32 is gathered into the existing INDEX16 consumer ABI. The helper scans every stride record, using both Uint8Array byteOffset and POSITIONT attribute offset. It can reject unused records conservatively; indices do not bypass scanning. Actual independent INDEX16 draws with backing byteOffset3, POSITIONT offset4, stride44 and unused first record pass; corrupting that unreferenced record rejects before allocation. The inherited structural fixture also checks the final record and partial stride. Existing consumer validation still rejects malformed ranges/counts/attributes.

All numerical fallback happens before guest_alloc(288) or native compilation. The ten-control counters prove admitted draws still allocate/compile; independent pixel tests additionally trap software fixedPrograms to disallow accidental fallback. Native descriptor origin/stage validation, bundle copy/free/finally, queue byte copying/order, separate guest/fixed lowering and backend selection remain as previously accepted. Ordinary tests freshly exercise their relevant copy/DEF/constant poisoning, selection and semantic rejecting controls. Retained five ownership/fault modes and released-IR overwrite evidence is reused, not misrepresented as real allocator reuse. Missing native export still falls back.

Moved-viewport correction is valid: vertices now lie inside the moved viewport, so narrowing admission does not turn the rebinding control into an invalid draw. Its changed native DEF/immutable prior IR assertions still run. A separate new moved viewport test verifies every expected covered pixel, not just changed bytes.

Executed evidence

node test/test-d3d9-shared-positiont-ir.js
node test/test-d3d9-positiont-numerical-admission.js
node --require ./scratch/runs/20261023-render-positiont-rhw-admission-repair/candidate-preload.js test/test-d3d9-positiont-numerical-admission.js
node scratch/runs/20261023-render-positiont-numerical-independent-review/independent-pixels.js geometry
node scratch/runs/20261023-render-positiont-numerical-independent-review/independent-pixels.js interpolation
node scratch/runs/20261023-render-positiont-numerical-independent-review/independent-pixels.js layout
node scratch/runs/20261023-render-positiont-numerical-independent-review/independent-pixels.js texture
node scratch/runs/20261023-render-positiont-numerical-independent-review/independent-pixels.js boundary-x
node scratch/runs/20261023-render-positiont-numerical-independent-review/independent-pixels.js boundary-y

All final commands exit0 except the pinned931aff producer substitution, which exits1 EXPECTED at finite reciprocal overflow ordinary route nonblack pixel: actual black versus expected red, before the new policy assertion. Its RHW1 control passes first. Candidate producer SHA and native inputs are pinned. Fresh numerical test: ten nonblack/10-red-pixel/full-frame controls and51 preallocation rejects pass. The exact finite RHW1e-39 repro now takes honest fallback and draws the established ten red pixels. Ordinary corrected depth cases, native selection, old MUL-to-ADD pixel rejection and DEF controls pass.

New independent script has28 admitted/full-frame cases:

Every final case requires non-null shared IR, actual native producer counters, successful draws on both paths, full-frame equality and an independent pixel or coverage oracle. Logs retain input IDs and framebuffer SHA256. All use the actual four production fragments,69097B module SHA256 5b12ae64af023d0203a5498ff594eaa38d9f7d90b50046533934f784e4c46ed8, exact8MiB fixture arena per process. No canonical/default compile occurred.

Intermediate fixture failures are retained: reversed winding initially moved z=1 to the origin but expected a LESS-depth pass; raw Uint32 was rejected by the existing INDEX16 interface; width2048 was initially attempted on8-wide target, then target2052 hit the existing2048 maximum. These are equal-path fixture errors, not production mismatches. Original scripts/logs remain, and final cases use valid targets and explicit depth. Texture expected values were strengthened from red diffuse masking to white diffuse/nonblack sampled color.