Independent POSITIONT source review, 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.

Verdict: REQUEST CHANGES for candidate 5eb16d60dd9f5302060509185fb2ebbbad88bba4, branch task/render-positiont-recovered-input-oct23, base bb9c496ed13095cc832fdf607f0937fb9c14b7de. Strict authenticated ls-remote independently matched candidate and main. Main is still the supplied base; this review performs no integration.

First correctness gap: POSITIONT depth and clipping regression

At lib/d3d9-software-backend.js:899, the migration disables the pretransformed raster flag for every shared native packet. The native fixed compiler's existing clip-space program reconstructs z as (POSITIONT.z - MinZ) / (MaxZ - MinZ); when the denominator is zero its scale is zero (src/09aj-d3d-fixed.wat:521). This program now enters homogeneous clipping and ordinary viewport depth projection.

The previous software POSITIONT path deliberately skips homogeneous clipping and copies screen x/y/z, retaining RHW for perspective varyings. See src/09ah-d3d-software.wat:584 and the explicit POSITIONT comment at line 683. The profile admits nondefault and equal viewport depth bounds without excluding the affected vertex depths. Therefore an otherwise supported draw changes observable pixels when it is migrated.

The independent depth-controls.js executes the actual native fragments and both production software routes. All cases use a white texture, red diffuse, triangle (0,0),(4,0),(0,4), RHW=1, LESS depth, and pixel (1,1):

The second and third rows are clipped away. The last two demonstrate both directions of the wrong depth decision: the shared path collapses depth to .5. Every case is accepted by the new producer and by the software viewport validator. This is an introduced software regression; no new GPU pixel regression is asserted (the old JS WebGL emitter already reconstructs z).

Hand production repair back to the coordinator: preserve the pretransformed depth/clipping semantics or narrow admission safely, then extend the normal pixel differential to these cases. Do not infer that shared IR bytes alone make raster semantics equivalent. No production file or candidate test was changed by this review.

Other reviewed boundaries and evidence

All ten donor checks independently replayed with exit 0. The exact focused module is 69,067 bytes, SHA256 49311866a815fdf9719c8cb4fa5d5f713fa4c189d2ea096d3d9a5558e6be5591, using the declared 8 MiB arena. This includes native texture/diffuse, depth, blend, unequal RHW, viewport DEF, guest-bank poisoning, four fallback image comparisons, eleven JS profile rejects, ten direct native descriptor rejects, and the MUL-to-ADD semantic rejecting control. Existing tests do pass; their viewport comparison uses z=.5 inside [.2,.8], so it misses this gap. No false pass of their stated individual pixel oracle was demonstrated.

ownership-controls.js independently verifies descriptor/bundle frees on success, zero descriptor allocation, zero native compilation, bad bundle projection, and a thrown postcompile memory read. It overwrites the actual released vertex IR storage and confirms the packet retains its bytes. The candidate's own storage-reuse assertion overwrites descriptor storage rather than native IR; this independent control supplies the stronger check. Native heap reuse is still unqualified: the arena's frees are no-ops. The new controls exit 0 when all expected rejection/cleanup assertions and all four demonstrated pixel differences hold.

GPU recording checks production shader selection/source/state and draw calls; it executes no GLSL and produces no actual WebGL pixels. Real WebGL1/2 compilation/pixels, full COM-to-bridge/canonical module and heap behavior, actual-driver readback counts/causes, games and performance remain unqualified. No full build, Boat, browser, runtime, lifecycle, deployment, worker or main write was performed. No test box is assigned. bx_fn7pn7yg remains owned by the coordinator; its state/expiry were not probed in this source-only task.

Reproduction and accounting

The measured whole allocation manifest was written before checkout: 26,757,246-byte upper bound, including donor/copy Git/source/tools, temporary and evidence allowances, and 8 MiB reserve; cap 67,108,864 bytes. Host free 6,726,496,256 bytes exceeded the 2,982,150,144-byte floor including projected growth. GIT_NO_LAZY_FETCH=1 applies throughout. A local isolated copy of the small focused donor supplied all inputs: no raw recovery, broad fetch/archive, or donor metadata change. All 57 materialized tracked candidate files match their Git blob identities and all 20 donor evidence seals verify.