skills/legaldesign/references/qa.md

# QA before delivery

Validate the specification, compare the source record, and inspect the rendered artifact. Schema success or a screenshot file is not proof of fidelity, legibility, or functioning interactions. Correct failures and rerun affected/dependent checks; disclose anything unavailable.

## 1. Intake and explanation

- Check the conversation actually paused after unanswered intake; an unanswered question followed by production fails the method, even if the resulting file works.
- Compare inputs → ledger → artifact and back. Preserve every material claim, qualification, uncertainty, option, and source status; justify omissions. Verify facts, question, qualified answer, supported relationship, and action under [method.md](method.md).
- Check one-page completeness or slide-overview coverage. On slides, visible facts/question/answer precede deep dives; each concise topic reaches its promised page exactly once. A topic must not also open a popup.
- Compare the chosen representation with its intended meaning. Verify comparison alignment, connections, membership, units, scales, and labels. Reject interchangeable cards that conceal a central dependency, but do not demand a diagram where text or a table explains better.
- For slide-brief issue pages, check the actual three-card analysis column and adjacent support column, not just the prose plan. Require a specific reason for a non-issue layout. The action must be sourced advice or labelled proposed follow-up, never a recommendation invented merely to fill the recipe. Source previews must match the recorded excerpt/clip and retain their source target after Save and both exports.
- Read every labelled arrow aloud as source → verb → target. Check that it expresses the supported direction: permission enables service; service depends on permission. Check negated outcomes too: preventing completion is not preventing blocked completion. A plausible label attached to the reverse arrow is still a wrong explanation.

## 2. Visual and responsive behavior

- Inspect every page in both themes at 1280×800, 1440×900, and 2560×1440, then resize back. Check the [fixed geometry/type contract](component-grammar.md): effective displayed desktop type must not enlarge or shrink with the window. Inspect transformed ancestors as well as computed font sizes. No clipping, overlap, empty navigation-caused pages, document scrolling, or rail/pager below the content frame.
- Inspect a cold first paint, including a Retina-scale laptop, before hovering or toggling a theme. Text must not snap into place only after interaction. Check headings, prose measure, card depth, figures, label clearance, and authentic source clips against the actual house components; geometry coordinates alone do not prove correct painted text.
- Inspect 430×932 and 390×844 phones: initial v4 load opens the existing 18px reader once with unchanged content; closing returns to the scaled canvas/book, and Read page reopens it. Check compact contents, responsive figures, and internally scrolling detail. Wide figures pan inside their frame rather than overflowing the reader. Any accordion fits expanded.
- Check light/dark semantic roles, subtle neutral reading feedback, visible keyboard focus/editor selection, explicit paint overrides, and final text contrast of at least 4.5:1, including configured firm colors. Dark popup prose should use primary ink, not metadata gray. A no-overflow pass cannot compensate for losing the explanation.

## 3. Reading and source interactions

- Open every overview topic and purposeful detail target with real pointer and keyboard input, including figure targets in phone reader mode. Forced/programmatic clicks do not establish usable hit areas; diagnose intercepting geometry or faulty targets instead of deleting the explanation.
- Confirm popup purpose, exact excerpt/clip correspondence, provenance/limits, and actual original link or honest unavailable state under [evidence.md](evidence.md). A working link or matching checksum does not establish verification. Supplied narrative and clip content must not be silently conflated.
- Test one-target semantics, short underlined text links, hover isolation, visible keyboard focus, Escape/scrim close, and return focus/reader position. No sticky pointer highlighting or competing navigation/popup action.
- Exercise overview topics, contents, Previous/Next, saved position, and compact phone navigation. Labels, destination, current position, and page count remain synchronized after reopen and export/template reset.
- For grouped overviews, open each group and every topic; check one-open behavior, expanded fit, keyboard access, and full coverage in working, client, template, and phone-reader modes. Export must reject overflow hidden in a closed group. Template group labels must be neutral placeholders.
- Cold-open grouped cards at 1800px and 2560px wide before any resize: each automatic card must fill its grid track. Check narrow-to-wide resize and preserve explicit editor widths across Save and both exports. Check the compact grouped index separately: manual collapse, keyboard activation, the correct active entry when group order differs from page order, automatic group reveal after Next/Previous, and no long flat list when subject groups exist.
- Pinch in and out repeatedly on phones, including reader/canvas, open detail, editing, and reopened exports. Displayed content must magnify, not shrink to compensate; pinch alone must not rerender a diagram or steal text-edit focus. Check a genuine resize separately and disclose simulated versus physical-device coverage.

## 4. Editor and persistence

- Exercise all editor capabilities in [component-grammar.md](component-grammar.md) with actual pointer/keyboard actions, not only state setters. Include HTML reflow and SVG geometry resizing, diagram labels, popup title/lede/sections, and undo/redo of added/duplicated objects and popup changes.
- Save and reopen the working HTML. Confirm text, geometry, paint, additions, popup edits, theme, page, and explicit decision response/note persist over the validated baseline. Source identity/status stay protected; no extra authoring utilities or unsupported comments layer appear.

## 5. Exports and privacy

- Reopen Export HTML: the single composition, intended reader interactions, navigation, spacing, user formatting, and permitted clip bytes work offline without editor chrome or authoring rationale. Test the post-export Open exported HTML preview separately from disk-save confirmation; canceled/failed export must not present a stale success link.
- Reopen Export template: reusable geometry, informative heading/navigation placeholders, and editing work. Scan visible/hidden markup, inert data, SVG, URLs, images, metadata, recovery identifiers, decisions, added text, and popup edits for private matter. Source clips become labelled neutral frames, with identifying bytes/metadata removed.
- Check strict serialization, allowlisted rendering, and absence of remote code/font/image dependencies. For package/runtime changes, regenerate and reopen every shipping template and run relevant repository validators/browser tests; a showcase screenshot or privacy scan alone is insufficient.

## Handoff

Give the working file, what it explains, checks actually run, and remaining input, source-access, or validation gaps. Do not claim a browser check from code inspection or a verified source from a functioning popup.