docs(skills): make activation recompute guidance evidence-based

[cuichenx]

Summary

• replace universal core_attn guidance with an architecture-, backend-, and peak-aware decision table covering every supported recompute label
• add matched H100 evidence for Moonlight 16B and Nemotron 3 Nano, with bounded historical Llama 3 70B context
• update the stable guide, skill metadata, and activation eval cases while clearly marking prior NVSkills results as historical

Why

The previous guidance generalized a qualitative module ranking from an unmatched Llama experiment. It did not account for architecture-specific boundaries, fused-attention behavior, nested boundaries, runtime compatibility, or the distinction between moving an OOM later and making a workload viable.

Evidence

Model	Evidence-based conclusion
Moonlight 16B	moe_act reduced peak allocation by 1.85% for a 1.99% step-time cost; mla_up_proj reduced it by 1.96% for a 5.19% cost; explicit core_attn produced no material memory reduction under fused attention.
Nemotron 3 Nano	Whole-moe was the smallest tested boundary that completed; adding layernorm recovered another 1.014 GB at a 3.07% step-time cost. Narrow moe_act saved 4.194 GB at the intermediate peak but still OOMed after advancing farther.
Llama 3 70B (historical)	Whole-mlp recovered about 3.2 GB but cost about 16% throughput in the prior unmatched experiment; this is retained only as contextual evidence.

The cross-model conclusion is to start with the narrowest architecture-specific boundary responsible for the measured per-rank peak, then broaden only when matched measurements show that it is necessary. There is no universal recompute-module ordering.

Validation

• uv run --no-sync pre-commit run --all-files
• skill package validation with quick_validate.py
• JSON and YAML parsing, Markdown link checks, privacy scan, and git diff --check
• independent pre-publication technical review
• all successful experimental rows reported finite losses with zero skipped or NaN iterations

Publication note

The existing skill.oms.sig belongs to the previous skill contents. This PR remains draft while a fresh NVSkills evaluation/signature refresh is requested; the historical signature must not be treated as validation of the amended package.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[cuichenx]

/nvskills-ci