Shared D3D9 POSITIONT native IR source candidate, 2026-10-23

Base: bb9c496ed13095cc832fdf607f0937fb9c14b7de. Branch: task/render-positiont-recovered-input-oct23. This implements one bounded source slice; it does not qualify the renderer or any game.

The supplied docs/render-command-unification.md proposal was read from the coordinator workspace. That path is absent from the exact base tree. Its Incremental implementation and Acceptance sections supplied the assignment; no copy of the proposal is introduced by this change.

The existing WAT fixed compiler already emits immutable native IR with DEF constant records. Software used that compiler, while WebGL used a separate JS fixed-function plan. WebGL's guest-IR entry point correctly refused the native fixed-origin flag. The migration therefore adds a bounded native export and a separate fixed-origin GPU entry point rather than relaxing guest validation.

Supported contract

The producer admits triangle list/strip/fan draws with no bound guest shaders, exactly FLOAT4 POSITIONT0, FLOAT4 or D3DCOLOR diffuse0, and FLOAT2/3/4 UV0 in distinct registers. Texture0 is a 2D texture; stage0 uses MODULATE with TEXTURE and DIFFUSE for both RGB and alpha. Stage1 terminates the cascade. Lighting, fog, specular, alpha test, color key, user clipping and outline fill are outside this profile. Ordinary existing depth/blend state remains in the draw packet. Other profiles and older native artifacts retain the previous fallback.

d3d_fixed_compile_positiont validates the bounded DFX1 profile and invokes the existing compiler. The read-only DFB1 adapter exports version1, positiont-modulate, d3d-clip, embedded-def, and copied vertex/pixel IR. The bridge frees the native bundle and descriptor before publishing the packet. Both executors consume these same serialized IR/DEF bytes, ignoring the guest programmable constant banks. Software bypasses its fixed-state compiler for the packet; WebGL bypasses its JS fixed plan.

Both use the clip-space vertex program, rather than selecting the old software-only screen-coordinate variant. Software therefore uses its ordinary clip-space raster projection. The GPU adapter retains the established D3D integer-center/top-left conversion of 0.5 - 1/16 pixel after IR execution. Queue transport, resource identities, leases, presentation, completion and readback code are unchanged. The recording GPU test observes zero readback or finish calls for the supported draw; these are test-double observations, not measurements of a browser driver.

Validation

node test/test-d3d9-shared-positiont-ir.js compiles only the four existing native IR/VM/raster/fixed fragments, unchanged except for the added export, with the vendored WATX compiler. It validates and executes a 69,067-byte module with an 8 MiB test arena. Module SHA256: 49311866a815fdf9719c8cb4fa5d5f713fa4c189d2ea096d3d9a5558e6be5591. The arena replaces guest allocation/address translation, does not reclaim allocations, and cannot qualify the canonical heap or its lifetime behavior.

Passing controls cover exact IR-byte differential against the existing native compiler; independently expected viewport DEF; immutable copies; changed viewport constants; poisoned guest constant banks; eleven unsupported-profile fallbacks and missing-export fallback; ten direct native descriptor rejects; independently expected native texture/diffuse pixels, depth order, alpha blend, and unequal-RHW interpolation; four full-image comparisons to the preserved native software fallback; and queued snapshot mutation behind an incomplete command. A MUL-to-ADD corruption executes but fails the same expected-pixel oracle, proving the semantic rejecting control.

The production GPU consumer is exercised with a recording GPU. The test makes both duplicate fixed compilers throw, confirms the native GLSL sources and raster adaptation, and observes a draw. It does not execute GLSL or produce GPU pixels. The existing position-format test is also run through the focused native module using the retained replay adapter; its assertions pass, but its canonical application/heap environment is not claimed.

Existing shader IR view, D3D9 shader lowering, fixed-plan cache, fixed top-left bias, GPU allocation/retirement, and neutral command-stream tests pass without the canonical build. Exact commands, exits and logs are in the evidence run. Test tier classification is automatic UNIT by the existing filename rules; no central test list edit is needed.

Outstanding acceptance

Real WebGL1 and WebGL2 compilation/pixels, complete guest COM-to-bridge drawing, canonical application compilation and heap lifetime, actual GPU readback counts/bytes/causes, and representative D3D/OpenGL/Glide game captures remain unexecuted. The five proposal Glide regression cases (NFS II SE, NFS III, Diablo II, Hitman and Hype), gameplay and performance are not qualified. No runtime, Boat, browser, lifecycle operation or deployment was performed. No test box was assigned; bx_fn7pn7yg remains coordinator/Shadow Man owned.

Evidence and recovery: scratch/runs/20261023-render-positiont-recovered-input/ contains admission, exact input recovery, measured dependency closure, commands/exits, logs and the existing-test replay adapter. The coordinator's recovered stream input is exactly 47,664 bytes, Git blob 1af9d8eb5f9cc476de30c9097452ae6870cf79cc, and is unchanged by this patch. All unmodified base paths are preserved in the thin index; publication uses write-tree --missing-ok because intentionally unmaterialized blobs are absent.