feat: clarify Before/After summaries and comparison navigation

[kentwelcome]

Summary

• Require concrete, parallel summary wording that distinguishes changed decisions/actions from changed explanations and presentation.
• Keep primary results and their full distributions visible alongside secondary summary leads in HTML and Markdown.
• Use one View behavior comparisons button; the primary-result box is context only. Unavailable evidence links to trial records.
• Derive counts from complete, unique named-trial assignments and retain memberships for auditing. Count-only raw extraction is replaced; persisted normalized reports remain readable without a schema bump.
• Preserve the new context in blinded quiz projections and update documentation.

Verification

• Decisions self-check, report-schema checks, live-report contract, hooks, and release-workflow checks passed.
• 13 quiz tests and 15 workflow tests passed; Ruff, Docker shfmt, and whitespace checks passed.
• Synthetic CLI rendering and desktop/mobile browser verification passed, including one-button navigation.
• Five approved Claude Sonnet re-extractions from saved evidence produced usable summaries with zero dropped rows. No new trials; original evidence hashes unchanged. All private artifacts remain outside the repository.

Independent review

Fresh read-only Standards and Spec subagents both APPROVE the final diff, including the single-link UX. No material findings or decisions required.

Limits

Membership validation proves bookkeeping, not semantic classification truth. Fresh extraction variation and unblinded previews do not establish improved human comprehension.

Refs DRC-4794.

[kentwelcome]

Fresh independent subagent review

Reviewed pinned committed diff ad143e5...c3ecb5b, following REVIEWER_GUIDELINES.md. Both reviews were read-only; no checks or model calls were rerun and no private evaluation artifacts were accessed.

Standards — APPROVE

No material findings, decisions required, or optional suggestions. Verified named-trial partition validation and caller cutover, derived counts, dropped-row reference remapping, saved normalized artifact compatibility, evidence-completeness gates, escaping, navigation targets, and blinded-context retention. No unauthorized mechanisms or privacy defects identified.

Spec — APPROVE

No material findings, decisions required, or optional suggestions. Concrete parallel summary instructions, visible primary-result distributions, mixed/unavailable safeguards, HTML/Markdown parity, and context-only result box with one shared comparison button match DRC-4794 and the latest UX decision. Quiz changes preserve methodology.

Scope and limits

Authorized additions: membership bookkeeping, derived result context, synthetic regressions, and documentation. Count-only raw extraction is removed; saved normalized evidence remains readable without a schema bump. Static correctness does not prove semantic classification truth or improved human comprehension.

Summary: 0 Standards findings; 0 Spec findings; no blockers.