Add an evidence-backed “Understand the change” report tab

[kentwelcome]

Summary

Add Understand the change, an evidence-backed explanation layer between the concise Summary and detailed comparisons.

• Keep all three Summary sections, including What this means.
• Place the Understand the change button under What the evidence shows.
• Open a separate top-level explanation tab; retain separate Behavior diff, Flow diff, Instruction changes, and Trial evidence tabs.
• Explain the distinction through annotated Before/After steps, practical meanings, unchanged behavior, limitations, full cited distributions including minority branches, and attributed source excerpts.
• Share explanation content across HTML and Markdown.

Evidence contract

The existing extraction call can supply a bounded, structured explanation. Validate comparison references, unchanged distributions, named source trials, and exact final-answer excerpts. Invalid optional explanations are omitted without discarding valid behavior comparisons. Rendering never calls a model.

Canonical report-data.json moves from schema 8 to 9 and explicitly records explanation availability. Saved run evidence without an explanation still renders with a transparent unavailable state and links to retained evidence; existing reports do not silently acquire new analysis.

All added fixtures are synthetic. No private prototype reports or real trial transcripts are committed.

Verification

• bash tests/hooks-test.sh — passed.
• python3 plugin/skills/behavior-diff/scripts/decisions.py --check — passed.
• bash tests/live-report-contract.sh — passed, including schema/explanation regression coverage.
• bash tests/release-workflow-test.sh — passed.
• Docker shfmt v3.14.0 check — passed.
• uvx ruff@0.16.5 format --check --diff . — passed.
• git diff --check — passed.
• Actual renderer/browser smoke: timing and mixed formula/writing cases, Summary button placement, explanation reload, source-excerpt deep links, separate evidence tabs, and 390px layout.
• Read-only load/render of an older saved case confirmed honest unavailable explanation, without re-extraction or modifying source evidence.

No live model execution was performed; synthetic rendering checks do not establish extractor adherence or improved reader comprehension.

Tracking

Closes https://linear.app/recce/issue/DRC-4810/add-evidence-backed-understand-the-change-report-tab

Independent review

• Specification review: APPROVE.
• Standards review: APPROVE after fixing a completeness-guard finding. Both HTML and Markdown now withhold stored conclusions when trial evidence is incomplete, blocked, invalid, or dropped; a failed-before/passed-after regression and browser smoke cover that boundary.