ci: refuse to publish a partial CUDA set (cuda12 coverage guard)

[danielhanchen]

What

Adds a publish-only guard to unsloth-prebuilt.yml that fails the run if the CUDA set about to be released is missing a runtime line. For linux and windows x64 it requires the cuda12 coverage classes to match the cuda13 ones before the draft release is flipped public.

Why

The latest published build (b9518) shipped the cuda13 x64 line but not cuda12, so every host with a cuda12 runtime falls back to the older b9493 (which also reports version: 1). I confirmed the fallback on real hardware across the fleet, all on cuda12 runtimes pulling b9493:

• T4 (sm_75), A100 (sm_80), RTX 3090 (sm_86), L4 and RTX 4090 (sm_89), H100 (sm_90), B200 (sm_100), RTX PRO 6000 and RTX 5090 (sm_120), DGX Spark (sm_121, arm64)

assemble_metadata.py builds the manifest from whatever app-* bundles are present with no completeness check, and the artifact download step is continue-on-error, so a partial set can still be assembled and published. This guard turns that into a hard failure instead of a silent manifest that strands the cuda12 majority.

Note on version: 1

Already handled on this branch by the fetch-depth: 0 checkout; the published b9493/b9518 assets just predate it. A fresh publish that includes cuda12 will also carry the correct build number.

Test

bash -n on the step plus a mock dist/: a complete set passes, dropping the cuda12 tarballs fails with the coverage-mismatch error, and the workflow YAML parses.