Backend portability: correct kernels and config on every CPU, wgpu and CUDA target #5338

Scope: Running correctly everywhere: CPU dispatch that picks a narrower kernel than the host supports, wgpu forward hardcoding rope_theta, platform arms for 0.73, orphan source files that never compile, and the pure-Rust cuda-oxide kernel path.

Why it matters: "Runs everywhere" fails when a target silently runs the wrong kernel or a wrong constant, and dark files hide code that no build checks.

Done when: Each target (x86 AVX2/AVX-512, ARM64 NEON, wgpu, CUDA) runs the model ladder smoke with output equal to the reference, and every .rs file under crates/ is compiled by some target.

Epic: #3999. Folded under triage rule 7b (2026-10-14): each row below was its own issue, closed as not planned and kept here as a sub-ticket. Reopen an issue when its work starts or its PR opens.

Sub-tickets

• Class: the shipped CPU path runs a narrower kernel than the host supports — dispatch returns before the wider ISA, kernels absent, module DCE'd, prefill run as decode #2880 Class: the shipped CPU path runs a narrower kernel than the host supports — dispatch returns before the wider ISA, kernels absent, module DCE'd, prefill run as decode (milestone 0.73.0); branch 1c/2880-vnni-reachable
  • accept: the outcome in the title, with a receipt
• OXIDE-001 (0.68.3, fast-track): cuda-oxide pure-Rust #[kernel] → PTX — one GDN kernel ported and measured on lambda-labs, gx10, yoga and mini; every kernel an ont:Kernel with parity, timing and safety receipts per silicon under pv lint --gate shapes #3522 OXIDE-001 (0.68.3, fast-track): cuda-oxide pure-Rust #[kernel] → PTX — one GDN kernel ported and measured on lambda-labs, gx10, yoga and mini; every kernel an ont:Kernel with parity, timing and safety receipts per silicon under pv lint --gate shapes (milestone 0.73.0); branch feat/3522-oxide-gated-rmsnorm
  • accept: the outcome in the title, with a receipt
• TR-16: 0.73 platform arms (TRACE-001) #4571 TR-16: 0.73 platform arms (TRACE-001) (milestone 0.73.0)
  • accept: the outcome in the title, with a receipt
• Every src file must compile: widen the #3809 dark-file guard to the whole tree (138 orphan .rs files, 53,211 lines) #4700 Every src file must compile: widen the 5 test files under apr-cli/serve/ (124 #[test] fns, incl. the apr-serve-v1/http-api-v1 FALSIFY contract tests) are never wired into mod.rs — never compiled, never run #3809 dark-file guard to the whole tree (138 orphan .rs files, 53,211 lines) (milestone 0.73.0); branch la-72/4700-dark-batch15
  • accept: the outcome in the title, with a receipt
• wgpu forward hardcodes rope_theta=1e6 (wgsl_forward.rs:970): Llama/Mistral wrong on wgpu (R12-1) #4832 wgpu forward hardcodes rope_theta=1e6 (wgsl_forward.rs:970): Llama/Mistral wrong on wgpu (R12-1) (milestone 0.73.0)
  • accept: the outcome in the title, with a receipt