Add quarto call axe, a hidden experimental site accessibility scanner

.claude/rules/dev-tools/axe-scan.md

@@ -0,0 +1,21 @@
+---
+paths:
+  - "src/command/call/axe/**"
+  - "tests/unit/axe-*.test.ts"
+  - "tests/smoke/axe/**"
+  - "tests/docs/axe-scan/**"
+---
+
+# Axe Scan Command (`quarto call axe`)
+
+For the scanner's architecture — pipeline, mode discovery, signatures,
+baseline semantics, exit codes, and the rationale behind each — see
+`llm-docs/axe-scan-architecture.md`.
+
+For how to use the command (flags, baseline workflow, CI recipe), see
+`dev-docs/axe-scan.md`.
+
+The signature normalization scheme is pinned by
+`tests/unit/axe-signature.test.ts`: changing `normalizeSelector` or
+`signatureOf` in a way that re-keys signatures requires bumping
+`kSignatureScheme` in `schemas.ts` on purpose.

.claude/rules/formats/html-dark-mode.md

@@ -0,0 +1,22 @@
+---
+paths:
+  - "src/format/html/format-html-info*"
+  - "src/format/html/format-html-bootstrap*"
+  - "src/command/render/pandoc-html*"
+  - "src/resources/formats/html/templates/quarto-html-*body.ejs"
+  - "src/core/brand/**"
+  - "src/command/dev-call/axe/**"
+---
+
+# HTML Dark Mode
+
+`formatDarkMode(format)` (`src/format/html/format-html-info.ts`) is the single
+predicate for "does this page have a dark mode" — `undefined` means no.
+The `data-mode` link attribute measures compiled CSS darkness; it does not
+declare the author's slot. The dark slot's stable identity is the
+`quarto-color-alternate` class.
+
+For the full picture — configuration surface (`theme`, `brand`/`_brand.yml`),
+rendered DOM markers, programmatic mode switching, and the known traps
+(light-only brand, light-colored dark slots, `theme: darkly`) — see
+`llm-docs/html-dark-mode-architecture.md`.

dev-docs/axe-scan.md

@@ -0,0 +1,192 @@
+# Scanning a site with `quarto call axe`
+
+`quarto call axe` scans a rendered Quarto site for accessibility
+violations with axe-core. It drives headless Chrome over every page, at
+desktop and mobile widths, in each colour mode the page ships. It groups
+violations by root cause, compares them against a committed baseline, and
+writes a report you can read, commit review comments from, or gate CI on.
+
+**Status: experimental.** The command is hidden (it does not show in
+`quarto call` help) and makes no stability promise — flags, artifact shapes,
+and semantics can change between prereleases. This page is contributor-facing
+documentation; it graduates to quarto.org when the command is unhidden (the
+same path the `axe:` render option's docs took).
+
+Prerequisite: a Chromium the scanner can find. `quarto install
+chrome-headless-shell` is the reliable route; an installed system Chrome or
+Edge also works.
+
+## When to reach for it
+
+The render-time `axe:` option checks the page you are looking at, in your
+browser, while you author. This command checks the *site*: every page, both
+viewports, light and dark, at audit time or in CI. Reach for it when you
+want a site-wide inventory, a regression gate, or output an agent can work
+through.
+
+## First scan
+
+```sh
+quarto render
+quarto call axe _site
+```
+
+The scan prints its matrix up front (`43 pages (40 light+dark, 3 default) ×
+2 viewports — 166 cells`), then one line per cell, then a findings summary.
+Artifacts land at the project root (the nearest `_quarto.yml` at or above
+the site dir):
+
+- `_axe-checks/findings.json` — every finding, machine-readable.
+- `_axe-checks/report.md` — the human summary, GitHub-flavored markdown.
+- `_axe-checks/README.md` — generated docs for the artifacts, including the
+  baseline how-to. Read this after a scan; it is written for whoever (or
+  whatever) has only the artifact directory in front of them.
+- `_axe-checks/cells/` — raw axe output per page × viewport × mode cell.
+
+`_axe-checks/` ignores itself (`.gitignore` with `*`): it is a disposable
+snapshot, not a thing to commit.
+
+### Reading report.md
+
+The report is GitHub-flavored markdown, made to be read where markdown
+already renders: your editor's preview, GitHub, or the sticky PR comment
+from the CI recipe below. Rendering it is optional. For a standalone HTML
+view:
+
+```sh
+quarto render _axe-checks/report.md
+```
+
+then open `_axe-checks/report.html` in your browser. The output lands
+beside the report, inside the self-ignoring artifact directory.
+
+`quarto preview _axe-checks/report.md` does **not** work from inside a
+project: `_`-prefixed directories are not project inputs, so preview stops
+with `No output created by quarto render report.md`. Use render-then-open,
+or pass `--report` a path inside your site source (see the flags table) to
+render and preview the report as part of the site.
+
+A finding is one *root cause*, not one element: an alt-less image in a
+shared include shows up as one finding with an instance count, not once per
+page. Grouping keys on axe's element selector (normalized), so this holds
+when the pages describe the element the same way — reliably true for
+Quarto's own chrome and for repeated template output, and occasionally
+wrong for an anonymous element whose surrounding DOM differs page to page
+(axe then picks different selectors, and one cause splits into two
+findings). Findings on many pages usually come from a shared source — a
+template, the theme, Quarto's own chrome — and one fix clears them all.
+
+## Flags
+
+| flag | default | |
+|---|---|---|
+| `--pages <globs>` | all `*.html` | comma-separated site-relative globs |
+| `--exclude <globs>` | — | skip globs, applied after `--pages` |
+| `--max-pages <n>` | ∞ | deterministic cap (sorted, first n) |
+| `--viewports <WxH,...>` | `1440x900,320x568` | |
+| `--themes <light,dark>` | `light,dark` | filters two-mode pages; one-mode pages always scan once |
+| `--timeout <ms>` | `30000` | per-cell budget |
+| `--settle <ms>` | `50` | extra delay after the page reports ready |
+| `--fail-on <impact>` | off | exit 1 on new findings at/above `minor`/`moderate`/`serious`/`critical` |
+| `--report <path>` | `_axe-checks/report.md` | put the report elsewhere, e.g. inside your site source |
+
+The narrow default viewport is 320 CSS px — the width WCAG's reflow
+criterion (SC 1.4.10) names, equivalent to 400% zoom on a 1280 px window —
+so every rule runs against the reflowed mobile layout.
+
+A subset scan (`--pages`, `--exclude`, `--max-pages`) says so loudly in
+every artifact: its counts describe the subset, not the site.
+
+## Running it as a post-render step
+
+The command works as a project `post-render` script, so a full render scans
+itself:
+
+```yaml
+project:
+  type: website
+  post-render:
+    - quarto call axe _site/
+```
+
+The script runs from the project directory after the outputs are written,
+so the relative site dir and the artifact anchor resolve exactly as they do
+on the command line. Exit codes propagate: without `--fail-on`, findings
+never fail the render; with `--fail-on`, a new finding at the threshold —
+or an incomplete scan — fails `quarto render` itself, with the scan's error
+line in the render output. This runs on every full render of the project,
+which adds the scan's runtime to each render.
+
+## The baseline workflow
+
+The first scan of a real site reports findings you will not fix today:
+upstream defects, deferred best-practice items, known false positives. The
+baseline is the committed ledger of those decisions — `_axe-baseline.json`
+at the project root, beside (not inside) `_axe-checks/`.
+
+1. Scan, then fix what you can from `report.md`.
+2. For each finding you are accepting instead: copy it out of
+   `findings.json` into the baseline's `findings` array, trim it, and write
+   a `note` saying why. The generated `_axe-checks/README.md` documents the
+   entry shape and the scoping rules (`pages: []` accepts site-wide;
+   a listed `pages` re-alerts anywhere else).
+3. Commit `_axe-baseline.json`.
+
+From then on, reports separate **new** findings from **baselined** ones,
+and only new findings can fail CI. A baselined finding that escalates in
+impact, or shows up on a page outside its scope, re-alerts as new. Entries
+a full-site scan no longer sees are reported as stale — prune them by hand.
+
+There is deliberately no `--update-baseline`: every entry exists because
+someone wrote it and said why.
+
+## Exit codes and the CI recipe
+
+| exit | meaning |
+|---|---|
+| `0` | scan complete; no new findings at/above `--fail-on` (when given) |
+| `1` | complete scan, new findings at/above the `--fail-on` threshold |
+| `2` | scan incomplete — a cell timed out or errored, no browser, nothing to scan. Takes precedence over 1: an incomplete scan never reads as a pass |
+
+A minimal GitHub Actions gate:
+
+```yaml
+- uses: quarto-dev/quarto-actions/setup@v2
+- run: quarto install chrome-headless-shell --no-prompt
+- run: quarto render
+- run: quarto call axe _site --fail-on serious
+```
+
+`report.md` is GitHub-flavored markdown, so posting it as a PR comment is
+workflow configuration, not tooling. A sticky comment beats a plain one:
+it updates in place on each push instead of stacking a new comment per run:
+
+```yaml
+- if: always()
+  uses: marocchino/sticky-pull-request-comment@v2
+  with:
+    header: axe
+    path: _axe-checks/report.md
+```
+
+(`if: always()` keeps the comment current when `--fail-on` fails the job;
+`header` keys the comment so other sticky comments are untouched. GitHub
+caps comment bodies at 65,536 characters; a whole-site report on a large
+site can exceed it — trim or attach as an artifact instead.)
+
+## Reading the output as an agent
+
+Point the agent at `_axe-checks/` and let it read the generated `README.md`
+first. Finding ids are stable across runs (`image-alt-6e3b76`), so "fix
+`image-alt-6e3b76`" is a well-defined instruction, and each finding's
+`occurrences[]` carries real selectors and HTML excerpts. Fixes belong in
+Quarto *source* (`.qmd`, `_quarto.yml`, `_brand.yml`, theme `.scss`) — never
+in the rendered site directory.
+
+## Where the pieces are documented
+
+- How it works and why: `llm-docs/axe-scan-architecture.md`.
+- What the artifacts mean, baseline entry shape: the generated
+  `_axe-checks/README.md` (regenerated every scan, always matches the build
+  that wrote it).
+- Render-time single-page checking: the `axe:` HTML format option.