feat: publish explicit CPU weight variants
Kernel-ready CPU files with explicit body encodings: FP32 for exact, INT16 group-64 for near-exact, and GPTQ INT8 group-64 for fast. The encoding-specific integer filenames replace the old near-exact/fast filenames; mode flags keep their meanings.
FP32 preserves the pinned checkpoint's parameter bytes and adds the precomputed vocabulary screen under a distinct format tag. The integer files are fresh repacks of the published files, preserving their tensor names, shapes, types, bytes and metadata values, including the original GPTQ overlay provenance. GPTQ was not rerun. No new accuracy or speed claim is made.
The model card describes the encodings, migration and rebuild prerequisites. The draft head, tokenizer, configuration and license are retained.
Validation covers direct content comparison, packed round trips, available model-backed gates, matching smoke/representative-page outputs, and full PR-bundle download/readback with smoke generation in every mode.
This catalog requires the matching runner changes tracked in GitHub PR #14. Merge that PR first, then merge this PR manually. Both merges are the user's steps; this PR does not change the Hub main branch until merged.
Merging as the PR on GitHub is done