2026-08-20 - Two wavefront models, agreeing

Done. Stubs for all three fidelity levels, then the second one built.

• Fidelity as a field of Strategy::Predicated, so invalid pairings are unrepresentable rather than rejected.
• wavefront.rs: registers as arrays indexed by lane, the mask in the scalar file where the hardware puts it, masked writes by select rather than branch.
• A differential test running four programs at both levels and comparing.

Verified. Gate green, device tests executed. Nine execution tests pass on hardware, including the two models agreeing on every program tried.

Surprises.

• The first attempt at splitting a long function saved one line. wavefront::new tripped too_many_lines; passing nine identifiers to helpers in a struct moved the length from the body to the call site and achieved 119 to 118. The identifiers were the bulk, so the helpers now allocate their own and return only what the caller uses. Worth remembering: when a split does not shrink anything, the split is wrong rather than the limit.

• Implementing a level broke the test asserting it did not exist, exactly as intended, and it is worth having felt: a loudly-stubbed path is pinned by a test, so building it fails the suite until somebody consciously removes the pin. That is the cheapest possible reminder to update the docs claiming it is a stub.

• Every lane must be switched on at entry. The mask is null-initialised, so left alone it is zero - every masked write would be discarded and the shader would produce a plausible buffer of zeros while executing nothing whatsoever. There is a test for it now, because that failure looks exactly like a shader that ran correctly and did nothing.

Not done. Subgroup is still a stub. The two built levels agree on four programs using four instructions, which is a start and not a proof. And the duplication between predicated.rs and wavefront.rs is real - deliberately deferred until two implementations existed, and now they do, so the seam is observable and the factoring is owed.