feat: guided report story and human-readable trial summaries

[kentwelcome]

Summary

• Implement approved guided-story A with five permanent report tabs, instruction-block highlighting, and explicit unavailable evidence states.
• Align each numbered trial Before-left/After-right with shared detail controls, visible final answers, and separately spaced trial links. Preserve exact command-path grouping and complete print output.
• Add short, human-readable per-trial summaries through the existing extraction call: takeaway, one sentence per side, and optional caveat. Validate exact record identities and retain raw evidence when summaries are missing or invalid.
• Keep HTML/Markdown information aligned and update README, architecture, and the approved decision record.

Data contract

Report data advances to schema v8 for typed decisions.trial_summaries. Regenerate reports from original run artifacts; older serialized report-data files are rejected. Existing decisions.json without summaries remains renderable with explicit unavailability. Rendering never calls a model. No new dependency or additional model-call stage.

Verification

Passed locally before commit:

• bash tests/hooks-test.sh
• python3 plugin/skills/behavior-diff/scripts/decisions.py --check
• python3 tests/report-schema-test.py
• bash tests/live-report-contract.sh
• bash tests/release-workflow-test.sh
• shfmt v3.14.0 check; Ruff 0.16.5 format check; git diff --check

Generated all synthetic gallery scenarios through the real ingestion/rendering pipeline. Browser smoke covered desktop/mobile trial alignment and summaries, grouped navigation, instruction highlighting, hover/keyboard/touch legends, mixed/blocked/missing/self-reported/planned-action evidence, and print disclosure-state restoration.

Two independent read-only reviews returned APPROVE with no material findings on both standards/correctness and spec/acceptance. Live-generated summary prose quality remains untested: verification used synthetic evidence, with no live model calls.

Tracking

DRC-4793 — Implement the approved guided-story Behavior Diff report

DCO-signed commit. No version bump or release included.

[kentwelcome]

Independent review results

Two independent read-only subagents reviewed the implementation now included in this PR (commit f5fbce1).

Standards and correctness — APPROVE

0 material findings. Reviewed the extraction → validation → loading → schema → HTML/Markdown pipeline. Exact trial identities, malformed and duplicate summaries, unequal sides, missing-summary states, and schema-v8 serialization preserve the underlying evidence. Escaping was checked in its actual HTML and Markdown output contexts. No new dependencies or additional model-call path were introduced.

Spec and acceptance — APPROVE

0 material findings. The implementation provides one short takeaway, Before/After descriptions, and an optional concrete caveat for each trial group. Summaries remain specific to that group rather than borrowing aggregate conclusions. The served mixed-case Markdown correctly distinguishes changed trials 1–2 from unchanged trial 3. Planned-action examples do not claim execution. Full answers and supporting evidence remain available, and documentation matches the implemented contract.

Review limits

These reviews inspected source, tests, synthetic fixtures, and served static content. The reviewers did not run tests, interactive browser checks, formatters, or live models, and made no edits. The implementation author’s prior deterministic and browser verification is documented separately in the PR description. Live-generated summary prose quality remains unvalidated; existing verification used synthetic evidence.

Summary: 0 standards findings, 0 spec findings; both reviewers approved.