SUN T1: prove CPU polynomial mutants survive worker smoke but fail semantic oracle

[isomorphisms]

SUN T1 — bounded semantic-test witness

Starts from Flexible Pipes #69 and AICI #183. Do not infer full shader or target-device acceptance from this change.

Original false-positive surface (preserved): tests/test_holomorphic_walk.c proves worker startup, finite score and normalized direction, but no mathematical field values. The unchanged smoke test is rerun against each corrupted evaluator.

Independent witness: tests/test_holomorphic_direction_semantics.c invokes the actual static direction_at() from the compiled C source, using a separate binary64 C99-complex cpow oracle, a fixed analytic fifth-degree anchor, zero/complex/boundary cases and 128 seeded cases. The oracle checks delta_q(u) and d(delta_q)/du, not physical z parity or the CPU sample-domain policy.

Adversarial controls: tests/semantic_mutations.sh compiles and executes the historical worker smoke plus the new independent checker against one valid control and five invalid variants: half-scale, derivative degree, omitted fifth term, wrong real sign and reversed coefficients. For mutants, a compile error or missing output cannot count as a kill. Each expected mutant must first pass the old smoke and then fail by SEMANTIC-MISMATCH in the new test. Source-shape drift fails closed. Receipts include exact checked-out revision, mutant source digests, test identities and seed, and retain stdout.

CI checks out the exact PR head and runs the bounded mutation test. The old worker smoke remains unchanged for reproducibility.

Explicitly NOT VERIFIED / no implied claims

• Executed Idris-generated GLSL and decoded framebuffer/RGB parity are not tested here. Holomorphic Audit math-to-GPU lowering: live renderer is still handwritten GLSL, not generated from the math #55/Repair GPU-flow CI gates and run the random-holomorphic PowerVR sweep #37 and draft Generate the live Holomorphic fragment from checked Idris source #63 remain separate.
• Whether the CPU's fixed u=z/6 sampling is an intentional policy or an error is undecided. This test asserts only the polynomial evaluator's meaning for supplied u; it does not change motion algorithms.
• No proof of deployed FP execution adapter, independent cross-repository AICI verdict, physical PowerVR/MIRO A1 acceptance, or global generator repair.
• This PR is one counterexample and smoke-vs-semantic landmark, not completion of SUN T1.

Next: AICI independent receipt/oracle audit, actual generated shader diagnostic output, fleet follower propagation and required-stage enforcement.

Review required; do not merge as if the broader SUN T1 acceptance had passed.