[skills] docs: codify recipe tuning workflow

[yaoyu-33]

Summary

• codify a fit-first, then fill recipe-tuning loop: keep dense TP as low as capacity allows, use the largest legal EP candidate, and prefer architecture-specific selective recompute
• document Qwen3.5 GDN handling through gdn_norm_out, with full-layer recompute retained as a measured capacity fallback
• require post-fit W&B plus per-rank memory review before A/B testing lower TP or higher MBS at fixed GBS
• keep alltoall as the correctness baseline and treat HybridEP as a matched candidate on NVL8 BF16

The EP8-on-8-GPU and EP32-on-32-GPU guidance is explicitly scoped to legal expert counts and parallel meshes. EP is documented as sharding expert weights only, not dense or activation state.

Validation

• five updated skills pass the standard skill validator
• modified JSON/YAML artifacts parse successfully
• uv run --no-project --active pre-commit run --all-files
• git diff --check origin/main...HEAD

No training recipe or product code changed, and no performance experiment was rerun. Existing measured dates remain unchanged; cards record a separate 2026-08-25 guidance refresh.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yaoyu-33]

/ok to test 376a9cb

[yaoyu-33]

/ok to test 32019a2