docs(evals): document when a skill eval needs a no-skill baseline arm and where to route plugin-level claims

[kyle-sexton]

Refs #5255
Refs #3614

Owner decision on #5255 (2026-09-29): adopt now, docs only. No runner change.

Summary

Document, for skill evals/evals.json authors, when a baseline (no-skill) arm is required and where to route whole-plugin claims.

• Recommendation: adopt now, docs and routing only. Basis: claude plugin eval already runs a with/without arm at plugin level (gap-analysis row 1, "already-covered"); the gap is that skill evals/evals.json files say nothing about when a baseline arm is required or where to route whole-plugin claims. Cheapest item, no runner change.
• Unblocks: authors of skill evals.json files get one rule for "when to use /evals:plugin-eval". Items 2 and 3 of evals: operator gate for the adopt proposals from the fleet-vs-Agent-Skills gap analysis (#3614) #5255 wait on this one.

Condition

Scope is documentation and routing only. Any runner or script change is out of scope and needs a separate owner decision.

Related

• evals: operator gate for the adopt proposals from the fleet-vs-Agent-Skills gap analysis (#3614) #5255 (operator gate), Evaluation-methodology vertical: gap analysis of fleet evals against the Agent Skills eval pipeline #3614 (gap analysis)