D553 - The translator draws

decided - 2026-09-04

Phase 6 step (c) - emit Fragment modules with output variables - and with it the export. exp is out of BLOCKED, which is now empty.

The first end-to-end check this project has been able to make

Phase 6's own page states the cost of building ahead of the spine: everything in it is verified against material this project generated, so it cannot be wrong in any way its own tests would notice. The framebuffer oracle was built to break that (D549, D550), and this is the first thing to go through it.

tests/translated_export.rs draws twice over the same red clear - once with the hand-assembled constant_colour_fragment_module, once with a module the translator produced from guest instruction words - and compares every pixel of the two frames. Neither side checks itself.

It is a differential, which is the shape every later comparison in this phase takes. The frames are compared and the expected bytes asserted, so a failure says which of the two moved.

What the fragment stage is

Three differences from a compute module and nothing else: the execution model, OriginUpperLeft, and whether the epilogue writes the observation window.

That last one is not cosmetic. A fragment shader may write a storage buffer only where fragmentStoresAndAtomics is enabled, and this project's device requests no features at all (D552). Skipping the epilogue keeps a fragment module inside what the device can run - and it also removes the only static use of the observation buffer, so the pipeline needs no descriptor for it. It costs nothing, because a fragment module's oracle is the attachment.

What the export translates, and what it refuses

Four sources read for lane zero, bitcast rather than converted - a vector register holds the colour's bit pattern already, and an arithmetic conversion would scale every channel - assembled into a vec4 and stored to the output.

Three refusals, each separate and each for its own reason:

• A module with no colour output. That is what Model::colour_output defaulting to None is for: a compute dispatch has nowhere to export to, and the failure that matters is not an error but a store somewhere arbitrary, which renders a frame that looks plausible and is not.
• Any target but mrt0. Which attachment another index selects is register state a guest writes, needs a capture, and D104 refuses to invent. Translating them all onto one attachment would appear to work.
• The compressed and done flags and the write mask. They live in the instruction's first word, are not among the operands the decoder solved, and are not read - so nothing here may claim to honour them.

orbistoun-spirv gained OpCompositeConstruct and its shape row; OpConstantComposite builds an aggregate from constants and an export's components are values.

Three breaks, and the one that did not fire is the useful one

• Assembling the components in reverse fails the frame comparison - [255,255,0,0] against [0,0,255,255], every pixel.
• Accepting a different target than the shader exports to fails at translation, with the refusal quoted.
• Reading lane one instead of lane zero passes.

The third is not a gap in the break, it is a fact about the shader: its registers are set by literal moves, and a literal move writes every active lane, so every lane holds the same value and the choice is unobservable. Distinguishing them needs a shader whose lanes differ, which needs an input that varies across a fragment - interpolation, which is not built.

That is written into the test rather than left out. A break that does not fire is evidence about the test's reach, and reporting only the two that worked would have overstated it.

What this does not do

It does not make a real shader translate. The subject is four moves and an export. No interpolation, no fragment inputs, no blending, no depth, one attachment, one target. exp was listed as the instruction blocking more shaders than any other, and it is now unblocked for the narrowest possible case - which is the case the oracle can check, and the only one that could have been checked first.

BLOCKED is empty and is kept. The distinction it draws - between nobody has looked at this and this is waiting on a subsystem - is what stops a worklist sending effort at whichever refusal is most frequent, and the next instruction that needs it will need it for the same reason.