feat(verification): place the explain-the-number and rerunnable-evidence principle parts in measure, confirm, performance verify and plugin-eval #12149

Parent #7589 decides that pstack's principles are placed "as reasoned sentences in the phase-skill step that applies them" (D58, Q174, Q175). The owner approved this 2026-10-15 (poteto-review decision rows 58 and 105). The wording rule the owner accepted (Q175) is that each placed part states the reason and the test that shows it applies, rather than a bare rule, and names the standard or discipline it comes from.

This slice places these parts from Lauren Tan's explain-the-number and prove-it-works principles:

• 7c: an untuned arm, left on default settings, is an alternative explanation for a measured gap. /performance:verify (plugins/performance/skills/verify/SKILL.md:119-124) names build modes and flags, but not an untuned arm.
• 7e: code-metric and quick performance compares state what limits the number and what was ruled out, or hand off to /performance:verify. /verification:measure (plugins/verification/skills/measure/context/performance.md:24) has neither line.
• 7f: before trusting an eval result, check that every run did the task, and that the gap holds across trials and models. /evals:plugin-eval's "Reading the delta" section (plugins/evals/skills/plugin-eval/SKILL.md:275) covers validity and trial noise, not a second model.
• 18e: script the check and keep output a reviewer can rerun. /verification:confirm's evidence manifest says to "cite a ## Reproduction block" (plugins/verification/skills/confirm/SKILL.md:218), not that the check be scripted.

Done when

• plugins/performance/skills/verify/SKILL.md's ruled-out list names an untuned arm (one left on default settings while the other was tuned) as an explanation to rule out, with its test.
• /verification:measure's compare output carries a Limiter: line and a Ruled out: line for code-metric and quick performance compares, or hands the compare to /performance:verify, and says which. The reason is that a number with no known limit cannot be trusted.
• plugins/evals/skills/plugin-eval/SKILL.md's reading step checks that every counted run did the task, and says that a gap claimed for more than the model it ran on needs a run on a second model.
• /verification:confirm's evidence step says that a check worth citing is scripted, with its command and output kept so a reviewer can rerun it, with the reason.
• Each added sentence states its reason and its test.
• .changes/performance/*.md, .changes/verification/*.md and .changes/evals/*.md fragments with bump: patch describe the additions.
• The PR's CI is green.

Parent: #7589