Clarify expensive-run exploration and sweep uncertainty

[wassname2]

Changes

• Keep the full anti-skimming form; distinguish it from the brief end message the user reads.
• Explain batching changes in expensive runs using distinct predicted effects, possible interactions, and conditional updates to the Bayesian mental model.
• Replace the sweep appendix's fixed t-statistic verdicts with raw paired comparisons and uncertainty. Distinguish exploratory decisions from claims of reliable superiority.

Targets dev4. Only SKILL.md and references/sweeps.md change.

Checks

git diff --check passed. A fresh-agent scenario check allowed batched changes, rejected artificial certainty from within-seed normalization, and retained the full form. It placed the form in the linked run log, so whether it must also appear in chat remains unresolved.

No actual training run or measured skill-uplift evaluation was performed.

Authored by Pi from wassname's requested changes.