# Release 14 — the social icons within reach, and a clear screen until the first Search

`main` at `ac438bbc` (release 12 merged and not rolled out; release 13 is a parallel session's).
Plan: [`export-header-first-search.md`](export-header-first-search.md), written 2026-09-28, with
slice HP added to it on the operator's ruling of the same day. Rules:
`plans/tools/implementer-rules.md`, with the commit trailer this release's prompts give.

**The operator's standing choices** (the plan, "The operator's words"; not re-opened):
- **No copy** beside any social link: icons with accessible names only.
- **A link the operator adds is an entry in `settings.json` `socialLinks`, never code.** The earlier
  tip-link branch is parked and not merged; nothing is taken from it.
- **No vendor file in the repository**, not even as a test fixture: the tracked tree is published
  by the source mirror.
- **Nothing is edited in the primary checkout**; each slice has its own worktree, and the parent
  merges with `git merge --no-ff` only on a clean tree.

## The slices

| Slice | Branch | What | Owns |
|---|---|---|---|
| HP | `homepage/social-visible` | The homepage's social row and one theme toggle in the header at every width, the wordmark's text dropped first on a very small screen and the row scrolling only as the last resort, one shared `SocialLinks` component, larger keys with a focus ring; Changelog in the footer only; the instance cards' names as the site's wordmark; the social icon checked by an allowlist on save and at render; a sized SVG with no viewBox gets one; `featured` ("Show in header") on a social link | `common/components/{SocialLinks,SocialScroll,ThemeRadios,ThemeToggle,ThemeScript,ThemeProvider,Wordmark}.tsx` + `themeConfig.ts`, `common/bin/doctor.ts`, the growth chart, `common/lib/{socialSvg,socialLinks}.ts` + tests, `common/lib/settingsSchema.ts` (the social-link type, parser, docs; the normalizer moved to `socialSvg.ts`), `common/lib/normalizeSocialSvg.test.ts`, `common/lib/{settings,site,homepage}.ts` (the save errors), `common/lib/{homepageSummary,siteColor}.ts`, `homepage/app/components/{Header,Footer,ArchiveCards}.tsx`, `homepage/app/lib/{nav,summary}.ts`, `homepage/app/not-found.tsx`, `homepage/e2e/**` (the fixtures, `helpers.ts`, the new and the rewritten specs), `homepage/playwright.config.ts`, `export/app/components/{MobileMenu,Footer}.tsx` (the `ThemeRadios` swap; the footer's read path), `editor/app/components/SocialLinksField.tsx` + `socialLinksJson{,.test}.ts`, `editor/app/{settings,sites}/actions.ts` (the save errors), `editor/e2e/settings.spec.ts`, `SETTINGS.md`, `SITE.md` |
| Lows, chart gap, T1, H1 (H3 folded in), H2 | `r14/two-grounds-headers` | The final review's Lows; the charts' surface gap; two grounds and each site in its own accent; the export and hub headers carry the social row and the toggle as one group, with an Archilyzer link to the homepage's `#instances` in place of the sites dropdown and the hub link; Changelog to the footer; after its review, the narrow header keeps the name and shows only the marked links (every header) | `common/components/{ThemeProvider,ThemeScript,ThemeToggle,SocialScroll}.tsx` + `themeConfig.ts` (and the deleted `ThemeMenu`, `ThemeRadios`), `common/components/charts/{ChartView,CrossSiteChart,surfaceGap}`, `common/styles/tokens.css`, `common/lib/{brand,accent,siteColor,paths,project,socialSvg,siteSchema,settingsSchema}.ts` + tests, `scripts/next-build-trace.test.mjs`, `export/app/components/{Header,MobileMenu,Footer}.tsx` (and the deleted `SiblingSwitcher`), `export/app/{layout.tsx,globals.css,changelog/page.tsx,lib/brand.ts}`, `export/e2e{,-hub}/**` (the theme, header and branding specs), `export/playwright.config.ts`, `editor/app/{layout.tsx,globals.css,sites/components/SiteForm.tsx}`, `editor/e2e/theme.spec.ts`, `homepage/app/{page.tsx,layout.tsx,globals.css,lib/*,changelog/page.tsx,components/{Header,ArchiveGrowthChart,ArchiveCards}.tsx}`, `homepage/e2e/**`, `homepage/content/docs/operate.md`, `SETTINGS.md`, `SITE.md` |
| HS | `r14/hidden-sites` | Hidden sites: `site.json` `listed` (absent = listed). An unlisted site builds and deploys as before, and is left off the homepage (cards, chart, `/stats`), the hub (members, federated search, `corpus.json`, `llms.txt`), every other site's footer and the published id lists (`channel-sites.json`, the pooled `stats/`); the channels only it exposes count in no public total. One checkbox in the site form | `common/lib/{siteSchema,site,homepageSummary}.ts` + tests, `common/controller/{poolSummary,buildStats}.ts` + tests, `common/bin/{compose-homepage,compose-hub}.ts` + `compose-hub.test.ts`, `editor/app/sites/{actions.ts,components/SiteForm.tsx}`, `editor/e2e/{helpers.ts,sites-crud.spec.ts}`, `homepage/e2e/{fixture-summary.ts,unlisted-site.spec.ts}`, `SITE.md`, `plans/FACTS.md` (Naming hazards) |
| S1 | `r14/first-search` | A clear screen until the first Search, on every site's search page and the hub's: no results area until the visitor asks (a Search, a profile load, or a link that carries a query or a filter), with the bar's "Press Enter or click Search to apply" line meanwhile; two page-life flags, the hold of release 8 unchanged and the gate new | `common/components/{SearchSessionContext,SearchResults,SearchBar}.tsx`, `export/e2e/first-search.spec.ts` (new), `export/e2e/{browse-all,workspace-shell,charts,restore-no-refire,responsive,tag-chips}.spec.ts`, `export/e2e/helpers.ts` (`showAll`), `export/e2e-hub/federated-search.spec.ts`, `export/CHANGELOG.md`, `plans/export-header-first-search.md` |
| CF | `r14/chart-fold` | The homepage's growth chart folds two or more sites each under 5 % of its total into one Other band on top, in its own near-neutral grey (`--chart-other`); the kept sites keep their colours; the legend shows them and Other, the hover titles and the table every site | `homepage/app/lib/growthGaps.ts` + test, `homepage/app/components/ArchiveGrowthChart.tsx`, `common/styles/tokens.css` (`--chart-other`), `homepage/e2e/{fixture-summary.ts,growth-chart.spec.ts,instance-colours.spec.ts,marketing.spec.ts}`, `homepage/CHANGELOG.md`, `plans/FACTS.md` |

**Order:** HP → `r14/two-grounds-headers` → HS (`r14/hidden-sites`) and S1 (`r14/first-search`),
siblings → CF (`r14/chart-fold`), off `main` after HS, with `main` merged in after S1. The shared
files are the three changelogs' `[Unreleased]` sections and this record.

## Record

### Slice HP, as shipped — the homepage shows the social links where they are seen (2026-09-28)

Branch `homepage/social-visible` off `main` `ac438bbc`, worktree `~/Projects/homepage-social-visible`
(block #3: editor 3301, test 3311, homepage e2e 3340, homepage static 3331), one Opus implementer.
Scratch files `hp-*` in the job's `tmp`. The rulings, all 2026-09-28, built in this order on one
branch:
1. The homepage's social links more visible; a change to the social-link schema where it has a
   reason.
2. The Changelog link moves to the footer; Base and Accent move behind a single options button with
   a gear icon that opens a modal.
3. The gear sits in the social icons' rhythm, with no gap of its own; the narrow header carries the
   icons and the gear in its bar.
4. On Official Instances, each site's name uses the bold-lead effect the sites' own headings use,
   with a slight tint or underline in the site's accent colour.
5. (The review, ruled by the parent.) No vendor file in the repository; the icon check hardened on
   save and at render; the shared dialog wrapper unchanged.
6. On the homepage chart, light mode has dark lines and dark mode has light lines.
7. On very small screens the header keeps its mark and drops the word; the social icons scroll only
   when they still cannot fit (rather than hiding icons on small screens).
8. The options menu is dropped for now in favour of a three-way toggle.
9. (The re-review, ruled by the parent.) A title or desc can no longer reach a page; an icon loads
   nothing from elsewhere; a stored icon the checker refuses does not block an unrelated save.
10. (After the merge, 2026-09-28.) The chart follows standard practice for light and dark grounds;
    the foreground-coloured separator lines are withdrawn.

**The branch's history was rewritten once** (after the review, ruling 5): its first nine commits,
one of which added a vendor file as a test fixture, were replaced by the five commits below,
re-committed from the same tree in the same logical steps, each tsc-clean. No vendor file is in
any commit of `main..HEAD`. Everything after `afc642fd` is new commits on top, and `main`
is merged in twice (`10cefd15`, then `918e5f85`).

**The Options dialog was built, then replaced by the toggle** (ruling 8): `OptionsDialog` (a gear
opening a modal with Base and Accent) shipped in `4d93dc84` and was deleted in `5da60518`, with
`options.spec.ts` and its screenshots. `ThemeRadios` stays: the export's slide-out menu renders it.

**What shipped.**
- **One row, `common/components/SocialLinks.tsx`.**
  - Props `links`, `placement: "header" | "footer"`, `className` (on the `<ul>`, which carries
    `data-social-links="<placement>"`); nothing site-specific.
  - Each link is a 36 px key (`size-9`) around the 20 px glyph, 44 px under a coarse pointer
    (`pointer-coarse:size-11`, Tailwind v4's own variant). The glyph is `text-muted-foreground`,
    `hover:text-foreground`, with a `hover:bg-muted` key, so a multi-colour icon has a hover state
    too. Focus is `focus-visible:ring-2 focus-visible:ring-ring` with
    `not-forced-colors:focus-visible:outline-none`: in forced colours a box-shadow is not drawn,
    and the browser's own outline is left in place. Nothing is drawn at rest.
  - `aria-label` and `title` are the label, `target="_blank" rel="noopener noreferrer"`, and the
    link has no text.
  - The header placement shows at most four (`headerSocialLinks`, below); the footer every link.
  - Every icon passes the save-time check AGAIN before it is inlined (`safeSocialSvg`, below). One
    that fails is not injected: the link shows its label as text.
  - The ids inside each inlined icon are scoped per row and per link (`scopeSvgIds`, below).
  - It uses only `useId`, so it works in a server or a client tree.
- **The header bound, `headerSocialLinks`** (`common/lib/socialLinks.ts`, pure). At most four: the
  links marked `featured` when any is marked, else all of them; of those, the last four. Marking is
  choosing: one marked of three shows one.
- **The homepage header** (`homepage/app/components/Header.tsx`): one group — the social row, then
  the theme toggle — in the bar at every width. The toggle is dressed as a key and its box sits
  directly after the last link's, so every glyph is 16 px from the next. From `md` (768 px) the bar
  is wordmark · nav · group, the nav's last link 32 px before the group's first box (40 px before
  its first glyph; the nav's own links are 24 px apart). Below `md` it is wordmark · group, and the
  four nav links have the rule below to themselves, where they fit at 320 px. No link is hidden by
  width (ruling 7): on a very small screen the wordmark's text goes first, and the row scrolls only
  as the last resort (see "The header layout").
- **Changelog is in the footer only** (ruling 2). `homepage/app/lib/nav.ts` declares `HEADER_NAV`
  (Docs, Source, Downloads, Stats) and `FOOTER_NAV` (the same, then Changelog). The footer's
  Sections and the 404 page read `FOOTER_NAV`.
- **One theme toggle** (ruling 8, replacing ruling 2's dialog): `common/components/ThemeToggle.tsx`
  with a new `variant="bare"` — dressed as a social key (36 px, 44 px under a coarse pointer, no
  border, the hover square, the ring, a 20 px glyph), cycling the base in `nextBase`'s order and
  named "Switch to {next}". Its default rendering, which the export and the editor use, is
  unchanged.
  - **The accent is pinned on the homepage**: with no accent control there, `ThemeScript` and
    `ThemeProvider` take `pinAccent`, so the stored `ytdlp-tb:accent` is neither read nor removed
    and the homepage keeps Signal, with no flash (the pre-paint script skips the read). Without the
    prop both are unchanged (unit-tested: the script string is identical).
- **The instance cards' names** (ruling 4).
  - `homepage-summary.json` `sites[]` gains an optional `wordmarkLead`: site.json's, resolved
    against `siteTitle` by `lib/brand.ts` `wordmarkLeadFor`, the resolver the sites' header and
    `siteSchema` use. It is still version 5: nothing reads the version to accept a file.
  - The homepage's loader keeps a lead only when it is a proper prefix of the title
    (`withCheckedLead`), so an older or hand-edited summary shows the plain title.
  - `ArchiveCards` sets the title with the shared `Wordmark` (lead 720, suffix 380) at the card
    title's size, and tints the lead in the site's own accent: `siteAccentColor`, `siteColor`'s
    accent half. A site with a lead and no accent keeps the foreground. `Wordmark` gains
    `leadStyle`.
  - `headerTitle` is not carried: the card's accessible name is `siteTitle`, and every live site
    has the two equal.
- **The social icon's SVG, checked by an allowlist** (`common/lib/socialSvg.ts`, pure;
  `settingsSchema.ts` re-exports `normalizeSocialSvg` and `socialSvgProblem`). The allowlist was
  chosen over the denylist: the element and attribute lists stay small, and every shape the
  existing tests and icons use passes.
  - The input is read tag by tag. It must be ONE well-formed `<svg>`: tags closed in order,
    attributes separated by HTML whitespace and quoted, no `<!…>`, CDATA or processing
    instruction. Comments, a leading XML declaration and a leading DOCTYPE with no internal subset
    are removed first.
  - Elements: shapes, groups, `defs`, `symbol`, `use`, gradients, `stop`, `pattern`, `clipPath`,
    `mask`, filters, `text`/`tspan`, and `animate`/`animateTransform`/`set`. No `script`, `style`,
    `foreignObject`, `a`, `image`, `title`, `desc` or HTML element (R1: a title or desc is an HTML
    integration point, where a child element left the icon unclosed around the page; one holding
    text only is removed first).
  - Attributes: the SVG presentation, geometry, filter and animation set, plus `aria-*`, `data-*`
    and `xmlns:*`. No event handler, whatever separates it.
  - Values: character references (numeric and named, with or without `;`) are decoded and the
    whitespace a browser ignores in a URL dropped before the checks. No `javascript:` or
    `vbscript:`; no backslash (a CSS escape) or CSS comment; no function that loads anything
    (`image-set(`, `-webkit-image-set(`, `image(`, `cross-fade(`, `element(`, `src(`, `paint(`,
    `@import`, `expression(`) — an animation's `to`/`from`/`values`/`by` included (R2).
    `href`/`xlink:href` only a plain `#id` as written; every `url(…)` only to `#id`, spelled so the
    render's scoping rewrites it (R6). A `style` holds presentation properties only (fill, stroke,
    stop-color, the opacities, stroke width/caps/joins, fill-rule, clip-rule, display, visibility,
    paint-order). An animation never targets anything named `href` or a handler (R5). Ids are
    plain names (`^[A-Za-z_][\w.:-]*$`).
  - The normalized OUTPUT is checked again, so no transform can assemble what the input check
    refused.
  - `socialSvgProblem` names the reason class. The editor's save error and the writers' errors
    append it ("… has an invalid SVG: it has an event handler attribute."); none echoes the markup.
  - `SETTINGS.md` and `SITE.md` (generated) say what an icon may contain.
- **The read path.** `safeSocialSvg` (`common/lib/socialLinks.ts`) runs the check again at render.
  `SocialLinks` and the export footer inline only what passes; a link whose icon fails shows its
  label as text, capped at 10rem with an ellipsis and the full label in `title` (R7). Every
  `dangerouslySetInnerHTML` of a social SVG goes through it (there are two). Each key clips its
  icon's paint (`overflow-hidden`, `contain: paint`; R4); the focus ring is the key's own shadow,
  outside that clip.
- **The write path** (R3): `socialLinksForSave` keeps a link whose `svg` is byte-identical to one
  already stored exactly as it is, and checks only a new or edited icon — in `writeSettings`,
  `writeSite`, `writeHomepageConfig` and the editor's settings and site actions. So a lane pause, a
  priority or a title never fails on an icon an older build stored; the render shows such an icon
  as its label, and `archilyzer doctor` names it ("social icons": file, label, reason class, never
  the markup). In this worktree the doctor line reads "3 icons in 1 file pass the check".
- **The size rule** (ruling 1): a root with no viewBox but a numeric `width` and `height`
  (unitless or px, decimals, either quote) gets `viewBox="0 0 W H"`, read from the root's own
  attributes (a size spelled inside another attribute's value does not count). A percentage,
  `em`, `auto`, a negative, or a missing or zero side is still refused, and a viewBox already
  there is never replaced.
- **`featured`**: optional on `SocialLink`, parsed only when exactly `true` and stored only when
  true, through `parseSocialLinks`, so `settings.json`, `site.json` and `homepage.json` all read
  it. The editor's `SocialLinksField` has a **Show in header** checkbox per row (Settings and a
  site's form), and beside it "With none checked, the header shows the last four." An old file
  with neither change parses byte-identically. `featured` also reaches each site's public
  `/site.json` (`buildSiteDescriptor` passes the link through): accepted, additive, and the hub
  does not render federated sites' links.

**The header layout** (built and screenshotted before each choice). Measured: the wordmark link is
148 px (152 px at 1× device scale), the four-link nav 276 px, a key 36 px (44 px under a coarse
pointer), the toggle the same.
- **No link is hidden by width** (ruling 7): the header shows its (at most four) links at every
  width. The bar is a size container (`@container/bar`), and when it is narrower than the full
  wordmark, a 12 px gap and the group need, the wordmark's TEXT is hidden and the mark stays (the
  link is still "Archilyzer home"). The threshold is a class per link count and pointer, in rem so
  it follows the text size (`WORDMARK_FITS`, measured, not guessed):

  | Links | Full wordmark with a mouse | Full wordmark with touch |
  |---|---|---|
  | 1 | from 276 px | from 292 px |
  | 2 | from 312 px | from 336 px |
  | 3 | from 348 px | from 380 px |
  | 4 | from 384 px | from 424 px |

  Measured at the threshold the gap is exactly 12 px; 1 px below, the text is hidden.
- **With the mark alone** (the fixture; the gap between the mark and the group): with a mouse, 3
  links — 320: 108 px, 340: 128; 4 links — 320: 72, 340: 92, 360: 112. With touch, 3 links — 320:
  76, 340: 96, 360: 116; 4 links — 320: 32, 340: 52, 360: 72. Four 44 px links, the toggle and the
  mark fit a 320 px screen.
- **The last resort** (`common/components/SocialScroll.tsx`): the row
  scrolls inside a box, its END shown first (the box is `direction: rtl`, the row `ltr` and
  `w-max`, so the first scroll position is the right edge — no script, no change to the DOM or tab
  order), no scrollbar, the toggle outside it, the header never wider than the screen. A link that
  takes focus is scrolled into view with its ring (the one bit of script: Chromium does not bring
  a partly hidden link into view inside an rtl box). From 300 px up, with default text, nothing
  scrolls; it takes 280 px with four 44 px links, or 320 px at a 200 % text size. Checked in
  Chromium and in an installed Firefox build (end first, nothing sideways); WebKit UNVERIFIED (the
  installed build does not match this Playwright).
- **What a visitor sees, by width** (default text size): the full wordmark from the table's width
  up; the mark alone below it, down to 300 px; the scroll fallback only below that, or at a much
  larger text size.
- **Two rows at 360 px** (a mouse, three links): the wordmark (20–168), 28 px, the group (196–340:
  three keys and the toggle), the 20 px gutter; below, the four nav links (20–291), 49 px to spare.
- **How it got here:**
  - First build, with two 32 px theme buttons and five nav links: three keys in the bar made a
    360 and a 390 px page scroll sideways by 36 px, so the keys were pinned at the end of the nav's
    rule, the nav in that rule below `lg`, and the row rendered twice.
  - Rulings 2 and 3 freed the bar: one gear, four links, and the group in the bar at every width.
  - The review found 320 px with a mouse scrolled sideways by 8 px; a width × pointer rule then hid
    links step by step (`6c500818`).
  - Ruling 7 replaced that rule with the collapsing wordmark and the scroll fallback (`0b3189fb`),
    and ruling 8 the gear with the toggle, the same size (`5da60518`).

**The cards' accent treatment** (ruling 4). Both treatments were built and screenshotted on the
three grounds at 390 and 1280 px, from a family-like summary: the six live titles, their site.json
leads (Jer, Hasan, Ani, Bonnell, Rekieta, Jaso) and their named accents; the numbers are synthetic.
- **Shipped: the tint.** The lead is in the site's accent, the suffix in the muted foreground.
  - The accent is one signal on the name and matches the card's stripe.
  - The underline sat under the card link's own hover underline, and on hover the lead drew two
    rules. An accent rule also reads as a link state.
- **Contrast of the lead against the card** (`--surface`): every named accent is at least 4.24:1 on
  Light, 4.21:1 on Sepia and 6.75:1 on Dark. A custom hex fitted to 4.5:1 on the ground is about
  4.1:1 on the card. The e2e checks ≥ 3:1 from the rendered colours.
- **The swap**, if asked: replace `{ color: accent }` in `ArchiveCards`' `SiteName` with a 2 px
  accent `text-decoration` on the lead. It is one line.

**The chart's separators** (ruling 6; after the review, a new commit on top). The line along each
stratum's upper edge in `homepage/app/components/ArchiveGrowthChart.tsx` was the page background
(`var(--background)`, 1.25 px): light on Light, dark on Dark. It is now `.growth-sep` in
`homepage/app/globals.css`: `var(--foreground)`, full strength, 1 px, and `CanvasText` in forced
colours. The gridlines are unchanged.
- **Three strengths built and screenshotted** (the foreground at 100 %, 60 % and 35 %, 1 px; the
  chart at 1280 and 390 px and a 4× close-up of the thin strata, on the three grounds; the
  published summary): `~/reports/release-14/shots/chart/sep-{100,60,35}-…`.
- **Shipped: 100 %.** Only full strength reaches 3:1 against any band; 60 % reaches it against
  none (at best 2.69:1) and 35 % against none (1.81:1). At 1 px, a 2–3 px stratum keeps a coloured
  stripe between its two lines in the close-ups; the old gap was 1.25 px.
- **Against the six bands, measured on the page** (the separator against each band's rendered
  colour):
  - Light: blue 4.25, amber 4.24, green 4.00, magenta 3.80, violet 2.17, rust 2.13 — minimum
    **2.13**;
  - Sepia: amber 3.55, blue 3.17, magenta 2.95, green 2.84, violet 1.81, rust 1.78 — minimum
    **1.78**;
  - Dark: rust 4.77, blue 4.63, magenta 3.64, violet 2.63, amber 2.54, green 2.49 — minimum
    **2.49**.
  With the ground's foreground, 3:1 is out of reach against those bands: Light violet and rust;
  Sepia green, magenta, violet and rust; Dark green, violet and amber. The series colours are
  unchanged, as ruled for this commit.
- **Homepage-only.** The separator is in the homepage's own chart. The shared charts
  (`common/components/charts/`, the homepage's `/stats`, the export sites and the hub) draw each
  series' edge in its own colour and have no page-coloured separator. Whether they take a
  foreground separator is left for the operator to rule on.
- **Withdrawn (ruling 10).** The separators were first drawn in the foreground colour, then
  withdrawn for a 2 px surface gap on the operator's ruling; that change is on
  `r14/two-grounds-headers` (its record, below).

**Beyond the prompt.**
- **Icon ids are scoped per copy (`scopeSvgIds`).** An id resolves to the first element carrying
  it, and a gradient defined inside a `display: none` copy does not paint. The first build rendered
  the header's row twice; with the ids unscoped, the gradient icon at 360 px lost its body and kept
  only its themed dot. The header's row is one element now, and the scoping stays: the footer
  inlines the same icons, and the export's header (H1) may render a row per breakpoint. Only plain
  ids are prefixed; references in `url(#…)`, `url('#…')`, `url(&quot;#…&quot;)`, `href` and
  `xlink:href` are rewritten, case-insensitively for the attribute names. A scoped, sized icon
  passes the checker again (unit-tested).
- **`sizeSocialSvg` moved** to `lib/socialLinks.ts` and is re-exported from `settingsSchema`.
- **The homepage e2e read the checkout's own `settings.json` and `homepage.json`.** The dev server
  now reads `e2e/.e2e-settings.json` through `SETTINGS_FILE` and an empty `e2e/.e2e-sites/` through
  `SITES_DIR` (both declared variables). The icons are synthetic: a gradient with one solid colour,
  one colour drawn white, and a two-colour disc with a letter pasted with only its size. Specs that
  need six links, none, or hostile icons rewrite the file (hostile ones raw, as a hand-edited file
  holds them), and an `afterEach` restores the trio. `marketing.spec.ts`'s "no link to ko-fi.com"
  runs against the fixture's links and still catches a link in code.
- **The footer's icons are `--muted-foreground`, no longer `--faint`**, the same as the header's.
  On the page ground that is 5.63 / 5.51 / 7.08:1 (Light / Sepia / Dark) against `--faint`'s
  3.45 / 3.59 / 3.53:1.

**Corrections to the prompt.**
- `NAV` had five entries at `ac438bbc`, not six.
- The homepage e2e baseline at `ac438bbc` was 36 passed, 0 skipped (1.0 min) in the empty-source
  state. The parent later copied a published source into the worktree, so the later runs are in
  the published-source state.

**What each icon looks like, per ground** (the fixture's synthetic icons, which behave as an
operator's pasted icons do):
- **The gradient icon:** the gradient keeps its colours on every ground. Its one solid part follows
  the link colour: slate on Light, brown on Sepia, warm grey on Dark, and the foreground on hover.
- **The one-colour icon:** the muted foreground on every ground (`#55646e` / `#6b5c43` /
  `#a39a86`), and the foreground on hover.
- **A two-colour disc with a dark outline and a dark offset disc:** on Light and Sepia the outline
  and the offset crescent show; on Dark both fall into the `#0c0a08` ground, and the disc reads flat.
  This is accepted, as ruled: no per-icon ring.
- **In forced colours:** single-colour parts take the link colour; the disc and the gradient keep
  theirs.

| sha | what |
|---|---|
| `518dd272` | `common:` `lib/socialSvg.ts`, the allowlist checker on save and at render with reasons; the size rule; `featured`; `lib/socialLinks.ts` (header rule, `safeSocialSvg`, id scoping, `sizeSocialSvg`); the writers' errors; the tests; `SETTINGS.md` / `SITE.md` |
| `a9dfa3de` | `editor:` "Show in header" with its hint; `socialLinksJson` + unit test; the save errors name the reason; the settings spec reads `featured` back |
| `4d93dc84` | `common:` `SocialLinks`, `OptionsDialog` (later deleted), `ThemeRadios`; `export:` `MobileMenu` renders `ThemeRadios`, the footer inlines only through `safeSocialSvg` |
| `6c500818` | `homepage:` the social row and the gear in the header at every width, the width × pointer rule (later replaced); Changelog in the footer only; the e2e's own settings and sites; the specs |
| `88d61908` | `homepage:` the cards' names as the site's wordmark, the lead tinted; `wordmarkLead` in the summary and its check on read; `siteAccentColor`; `Wordmark` `leadStyle`; `instance-wordmark.spec.ts` |
| `afc642fd` | `plans:` this record; the plan's slice HP, H1 and H2; STATE; the changelogs |
| `031c2115` | `homepage:` the growth chart's separators in the ground's foreground, 1 px, full strength (ruling 6); `growth-chart.spec.ts`; the record, the plan, the changelog |
| `0b3189fb` | `homepage:` the wordmark's text drops first on a very small screen (a container query per link count and pointer); `SocialScroll`, the last-resort scroll box, end first; no link hidden by width (ruling 7) |
| `5da60518` | `homepage:` one toggle (`ThemeToggle variant="bare"`) in place of the Options dialog; the homepage's accent pinned (`pinAccent`); `OptionsDialog` and `options.spec.ts` deleted; `toggle.spec.ts` (ruling 8) |
| `68ad1464` | `common:` the checker drops title/desc, refuses loading functions, escapes and comments, allows presentation styles only, plain references, reasons that say how to export; the review's vectors (re-review R1, R2, R5, R6, R8, R9) |
| `0be068cf` | `common, editor:` `socialLinksForSave` — an unchanged stored icon never blocks a save; `archilyzer doctor` names stored icons that fail (R3) |
| `4f7bc097` | `common, export, homepage:` each key clips its icon's paint; a refused icon's label bounded; `svg-vectors.spec.ts` (the parser invariant, nothing fetched); the hostile e2e checks no other origin (R1, R2, R4, R7) |
| `4d11542c` | merge `main` (`10cefd15`, the stats fix); `homepage/CHANGELOG.md` both sides kept |
| `fe9fb4b2` | merge `main` (`918e5f85`, slice Q's umtool build trace); `editor/CHANGELOG.md` both sides kept, `main`'s bullet under the one it refers to |
| _this_ | `plans:` this record; the plan's slices HP, H1, H2 and the new T1; STATE; the changelogs |

**Gates** (at `4d11542c`, after the first merge of `main`; logs `$T/hp11-*.log`. The second
merge, `fe9fb4b2`, brings only umtool, `scripts/`, plans and an editor changelog bullet; tsc and
`test:scripts` were run again after it: see the last line):
- **tsc** was clean before each commit since `afc642fd`, and before this one.
- **Unit:** common **2,202/2,202**; editor unit **87/87**; homepage unit **8/8**; `test:scripts`
  **185 + 1 skipped**; mcp **271/271**.
- **Docs:** `settings example --check`, `docs files --check` and `docs env --check` all exit **0**.
- **Doctor:** the new section prints `social icons` / `ok  stored icons  3 icons in 1 file pass the
  check` (the worktree's `settings.json`, the parent's copy; no other file there holds icons).
- **Builds:** homepage **ok** (22 s, `pnpm --filter homepage exec next build` with the fixture
  icons), export **ok** (34 s), editor **ok** (59 s).
- **Homepage e2e, full suite:** **75 passed, 0 failed, 0 skipped (1.7 min)**: `social` 24,
  `marketing` 10, `toggle` 9, `stats` 7, `source` 5, `brand` 4, `docs` 4, `downloads` 3, `theme` 3,
  `growth-chart` 2, and one each in `svg-vectors`, `instance-wordmark`, `instance-colours` and
  `no-data`. `options.spec.ts` went with the dialog.
- **Export e2e** (`site-branding brand related-sites theme theme-accent responsive archives-off`):
  **35 passed, 0 failed (1.6 min)**. The theme specs are unchanged: `ThemeToggle`'s default
  rendering is the export's as before.
- **Editor e2e** (`settings sites-crud`): **26 passed, 0 failed (1.8 min)**.
- **Browsers:** the scroll box's end-first start and focus scrolling are covered in Chromium by the
  suite and were checked by hand in Firefox at 240 px. **WebKit is unverified**: the installed
  revision does not match this Playwright's.
- **Earlier rounds:** at `88d61908`, common 2,179, editor unit 86, mcp 269, homepage e2e 68, export
  e2e 35, editor e2e 45 (with `export-search`); before the review, homepage e2e 36 → 50 → 59 → 63.
- **Screenshots** (`~/reports/release-14/shots/`, at 2×):
  - `hp4/` (64), the narrow header, from a dev server on 3330 reading synthetic settings
    (`SETTINGS_FILE`: the fixture's three icons, or six so the header shows four) and an empty
    `SITES_DIR`:
    - `header-{3,4}icons-{320,340,360,390}-{fine,coarse}-{light,sepia,dark}.png`;
    - `header-{3,4}icons-{768,1280}-{light,sepia,dark}.png`;
    - the wordmark is collapsed in every 320 and 340 px shot, at 360 px in all but three icons with
      a mouse, and at 390 px with four icons under touch;
    - `fallback-280-coarse-rest-dark.png`: four icons under touch at 280 px, the box's end in view;
    - `fallback-280-coarse-start-dark.png`: the same after Tab reaches the first icon, the box at
      its start;
    - `fallback-320-200pct-text-{rest,start}-dark.png`: the same pair at 320 px with the text at
      200 %.
  - `hp5/` (19), the toggle: `header-{320,360,390,768,1280}-{light,sepia,dark}.png` (three icons,
    the toggle last) and `forced-toggle-{unfocused,focused}-{light,dark}.png`.
  - `chart/`: the separators at 100, 60 and 35 %, per ground (the chart's section above).
  - `hp3/`: the cards' tint and underline; the header work does not reach them.
  - `hp2/` (the gear header) and `hp/` (the first round) are superseded; `hp2`'s dialog shots are
    deleted.
  - **The worktree's `homepage/out` is a fixture build: never deploy it; rebuild first.**
- **Numbers tool:** none.
- **After the second merge** (`fe9fb4b2`; `$T/hp12-post.log`): tsc clean in every package (42 s);
  `test:scripts` **188 + 1 skipped**, `main`'s new guard on umtool's build trace among them.

**They bite** (each change made by hand in the worktree, the specs run, the change reverted):
- The separators in `var(--background)`: `growth-chart.spec.ts`'s ground test fails (Light: the
  stroke is the ground's own colour).
- `WORDMARK_FITS` taken off the wordmark: 9 of `social.spec.ts` fail (320, 340 and 360 px with a
  mouse; 320–390 px under touch; the 280 px fallback; the 200 % text).
- `pinAccent` taken out of `layout.tsx`: `toggle.spec.ts`'s stored-accent test fails.
- `title` and `desc` back in the allowlist: `svg-vectors.spec.ts` fails on `title_child_el` (the
  page's `<main>` parsed inside the icon).
- The loading-function check skipped: the hostile-icon test fails (an `image-set(` icon is inlined
  instead of its label).
- From the earlier rounds:
  - `social.spec.ts` with the id scoping, the header bound and `pointer-coarse:size-11` removed:
    5 failed.
  - The hostile-icon spec fails without the read path.
  - The loader test fails without `withCheckedLead`.
  - `instance-wordmark.spec.ts`'s "Fix" + "ture Three" fails if the wordmark's spans stop being
    adjacent inline text.

#### Review (verdict SHIP AFTER FIXES; `$T/hp-review.md`)

| Finding | Fix |
|---|---|
| M1: 320 px with a mouse scrolled sideways by 8 px | `6c500818`: the width × pointer rule, since replaced by ruling 7 (`0b3189fb`) |
| M2: the `ui/dialog.tsx` change restyled the editor's command palette | `4d93dc84`: the wrapper is `main`'s, byte for byte (the dialog it served is deleted since) |
| M3: a vendor file as a test fixture would be published by the source mirror | ruled: no vendor file. The branch was rewritten so the file and its notice never entered it; D1 is tested with synthetic sized SVGs; the e2e disc is synthetic |
| M4: five inputs passed the normalizer and ran script | `518dd272` (the allowlist checker, the output re-check, the reasons), `4d93dc84` (the read path in both renderers); the review's five inputs and each class tested at the normalizer, the read path and in the browser |
| L1, L2: id scoping | `518dd272`: plain ids only (the checker refuses others; comments are removed), entity-quoted and case-varied references |
| L3: tests that could pass broken | the loader checks the lead; a mid-word lead; keys 36/44 px exactly; the glyph's own colour; six links under touch at 360; 320 px |
| L4: "below 390" vs `max-[389px]` | superseded with the width rule (ruling 7) |
| L5: `homepage.json` could win in the e2e | `SITES_DIR` → an empty directory. The reused dev server is pre-existing and left |
| L6: the dialog opened on the first radio | fixed in `4d93dc84`; the dialog is deleted since (ruling 8) |
| L7: `hp3` at 390 showed an old header | retaken |
| L8: `featured` in `/site.json` | accepted, recorded above |
| L9: the fixture build in `homepage/out` | recorded above: never deploy it |
| L10: the service's name in tracked files | gone from every file this branch adds or changes |
| L11: the size read from another attribute's value | the size comes from the root's own attributes |
| L12: the fixture echoed an operator icon's colours | the fixture's colours changed; the pre-existing test file is outside this branch |
| L13: `label` said "Visible name" | reworded; `SETTINGS.md` / `SITE.md` regenerated |

The answers, as ruled: the tint is kept; the footer colour `--muted-foreground` is kept; marking is
choosing, kept, and said beside the checkbox. The touch rule was kept, then superseded by ruling 7.

#### Re-review (verdict SHIP AFTER FIXES; `$T/hp-review.md`, "New findings")

| Finding | Fix |
|---|---|
| R1: a `title` or `desc` with an element child left the page's parser inside the icon | `68ad1464`: both are out of the allowlist; one holding text only is removed before the check, any other is refused as an element. `4f7bc097`: `svg-vectors.spec.ts` parses every accepted input as the static page carries it and asserts the page after it stays outside the icon |
| R2: an icon could load from another origin | `68ad1464`: `\` and `/*` refused in every value after decoding; `image-set(`, `image(`, `cross-fade(`, `element(`, `src(`, `paint(`, `@import` and `expression(` refused in every value; `style` holds presentation properties only (an allowlist); a `url(…)` count that differs raw and decoded is refused; the reviewer's eight inputs in the unit tests (`socialSvg.vectors.ts`). `4f7bc097`: the hostile e2e and `svg-vectors.spec.ts` fail on any request to another origin |
| R3: a stored icon the allowlist refuses blocked every save | `0be068cf`: `socialLinksForSave` checks only a new or edited icon, and keeps an unchanged stored one byte for byte, in the settings, site and homepage writers; `archilyzer doctor`'s "social icons" line names each failing icon by file and label, with the reason; the editor changelog's upgrade text |
| R4: an icon could paint and catch clicks outside its key | `4f7bc097`: each key, in the header and both footers, is `overflow-hidden` with `contain: paint`, so a `class` or `overflow` stays inside it; `68ad1464`'s style allowlist refuses `position`, `inset` and `z-index` |
| R5: an `attributeName` naming an href | `68ad1464`: an `attributeName` containing `href`, or starting `on` after any prefix, is refused |
| R6: an encoded or padded fragment passed the check but was not scoped | `68ad1464`: the raw value must be a plain fragment |
| R7: a refused icon's label had no width bound | `4f7bc097`: at most 10rem (`max-w-40`) with an ellipsis, the whole label as its title, in the header and both footers |
| R8: the comment on the echoed name | `68ad1464`: the echoed name is stripped to letters, digits and `_.:-`, at most 40 characters |
| R9: "remove the style block" drops the colours; a DOCTYPE's reason | `68ad1464`: the reason says to export with presentation attributes rather than a style block (in Inkscape, save as Plain SVG); a DOCTYPE with an internal subset has its own reason; the editor changelog says the same |
| R10: the branch moved during the review | nothing to fix: every commit after `afc642fd` is for the next review |

**Found and left:**
- **The export's header and the rest of its footer** are slices H1 and H2; **T1** ("two grounds")
  is planned, not built.
- **The shared charts' separators** (`/stats`, the export's charts) are still drawn in the page's
  background colour; ruling 6 names the homepage chart. A follow-up for the operator to rule on.
- **WebKit** is unverified for the scroll box (see Gates).
- **The Settings form's hint** still says the default links show "in every site's footer"; true of
  the sites until H1. Operator-facing; left alone.
- **`export-search.spec.ts`'s "a site's own social links win over the global default" depends on
  the order of the specs.** It reads `editor/test-settings.json`, which only a settings save
  creates, and fails when it runs first in a fresh worktree. The fix is a `.catch(() => ({}))` on
  the read, as `auto-queue.spec.ts` does.
- **`next dev` (16.2.3) refuses a second dev server in the same app directory.** A hand-started
  homepage dev server must be stopped before the suite runs; `reuseExistingServer` would otherwise
  reuse one started without the fixture's environment (pre-existing).
- **A pasted Illustrator or Inkscape file** with a `<style>` block, `<metadata>` or `inkscape:*`
  attributes is refused, with the element or attribute named and the export setting to use.
- **The hub's official cards** still show plain titles; ruling 4 names the homepage only.
- **At a 200 % text size the homepage's own content** is wider than a 320 px screen; the header is
  not (the e2e checks the header only).

**Changelog.**
- `homepage/CHANGELOG.md` `[Unreleased]`, worded as the end state, after `main`'s stats bullet: the
  icons in the header at every width beside the toggle (the wordmark's text drops first, the scroll
  box is the last resort); one theme toggle and the homepage's own accent; Changelog in the footer
  only; the cards' wordmark names; the chart's separators; the larger keys with a focus ring; the
  e2e's fixtures and specs.
- `editor/CHANGELOG.md` `[Unreleased]`, at the end of the list: the size rule and "Show in header";
  the icon check with reasons, on a new or edited icon and at render, and the upgrade note (an
  unchanged stored icon is kept; `archilyzer doctor` names those that fail).
- `export/CHANGELOG.md` `[Unreleased]` (created by `fix/stats-cache-key`): an icon that fails the
  check is shown as its label, bounded, and every icon paints inside its box.

### Branch `r14/two-grounds-headers` — the final review's Lows, the chart's gap, T1 and H1/H2 (2026-09-28)

Branch `r14/two-grounds-headers` off `main` `c6b8fc70` (slice HP merged), worktree
`~/Projects/homepage-social-visible` (block #3: export dev 3300, export e2e 3320, hub e2e 3341,
homepage e2e 3340, editor test 3311), one Opus implementer. Scratch files `t-*` in the job's
`tmp`. The rulings, all 2026-09-28:
1. (The chart, operator.) The chart follows standard practice for light and dark grounds; the
   foreground-coloured separator lines are withdrawn.
2. (T1, operator.) Sepia is dropped in every app, the editor included; a reader who chose Sepia
   gets Light; readers no longer pick an accent — each site shows its own, and stored accent
   choices are ignored.
3. (H1/H2, operator.) Every site's header and the hub's carry the social row with the cycling
   toggle as the last item of the same group, the homepage's; the Sites dropdown is replaced by a
   text link to the Archilyzer home's Official Instances (omitted on the hub); the header's Hub
   link is removed; Changelog joins the footer after Use with AI; on narrow widths the wordmark's
   text drops first and the scroll box is the last resort; the slide-out menu keeps the nav; the
   editor's header keeps the toggle, with no accent picker and nothing else restyled.
4. (The final review's Lows, parent.) F1, F3, F4, F5 and F8 fixed; F2, F6 and F7 left.
5. (After the review, operator.) On narrow screens a site's header keeps the site's NAME and shows
   only the link(s) marked for the header; the other social links are in the footer.

| sha | what |
|---|---|
| `d2b040fc` | `homepage:` the social row's scroll box and the compact nav are not tab stops of their own (F1); the toggle's no-flash test records every accent change (F5) |
| `ae5d535c` | `common:` the how-to-export hint only on a drawing program's leftovers (F3); stale text: `MobileMenu`'s comment, the older homepage changelog bullet, `featured`'s doc (F4) |
| `ec64011f` | `common, homepage, export, scripts:` every cwd-derived path op the three Next apps bundle opts out of Turbopack's tracing; the guard, renamed `scripts/next-build-trace.test.mjs`, covers them (F8) |
| `80228398` | `homepage, common:` chart bands are parted by a 2 px gap in the surface's colour; the foreground separators are withdrawn (ruling 1) |
| `3d0d1a46` | `common, export, editor, homepage:` two grounds, Light and Dark; each site wears its own accent (T1, ruling 2) |
| `46bd173c` | `homepage, common:` `id="instances"` on Official Instances and `INSTANCES_URL` (H1; the planned H3 folded in) |
| `04a1cfac` | `export:` the header's group, the Archilyzer link, no sites menu or hub link, Changelog in the footer, the narrow header; `header.spec.ts` (H1, H2) |
| `0f358ee7` | `export:` the footer's dot shows only after a downloads link; the social row keeps the shared row's inset |
| _this_ | `plans:` this record; the plan (H3 folded into H1, T1 and H1/H2 as built); STATE; FACTS' guard entry; the changelogs |

#### The final review's Lows

- **F1:** `SocialScroll`'s box has `tabIndex={-1}`, and so has the homepage's compact nav, which
  overflows below 320 px. Firefox 146 (the installed build, by its path) made both a tab stop of
  their own when they overflow. At 220 px with four links, without the fix Tab went home → the box
  → the icons; with it, home → the icons → the toggle → the nav's links. The suite is Chromium
  (Playwright's own Firefox revision is not installed), so it asserts the attribute.
- **F3:** the hint ("export it with presentation attributes rather than a style block (in
  Inkscape, save as Plain SVG)") follows a `style`, `<metadata>` and an element or attribute in an
  editor's namespace. An `<a>`, `<image>`, `<title>` with markup, `<foreignObject>` or `src` gets
  the reason alone. There is a unit test both ways.
- **F4:** `MobileMenu`'s comment, the older homepage changelog bullet that let readers pick an
  accent (reworded to the end state), and `featured`'s doc, which said "a narrow header shows
  fewer" (`SETTINGS.md` and `SITE.md` regenerated).
- **F5:** the stored-accent test installs a MutationObserver from an init script and records every
  value `data-accent` holds; a flash between two samples fails it.
- **F8:**
  - `/* turbopackIgnore: true */` is on the homepage's `source.ts` directory join and on the
    `docs.ts`, `snapshot.ts`, `summary.ts` and both changelog pages' cwd joins.
  - In `common/lib/paths.ts`, every join on a path built from the repo root goes through one
    opted-out `under()`, and `findMonorepoRoot()`'s walk and fallback are opted out.
    `getPaths()` returns the same 58 values as before (diffed).
  - The guard is `scripts/next-build-trace.test.mjs`, in `test:scripts`, 6 tests.
    - It scans umtool as before, and adds `homepage/app`, `export/app`, `editor/app`,
      `editor/lib`, `editor/instrumentation.ts` and every `common/` module but `bin/` (over 500
      modules).
    - A function declared in the file whose body carries a source is a source.
    - Brackets inside a regex literal no longer end a call.
    - On `main`'s files it lists 57 findings. With `main`'s `source.ts` put back, it fails on
      `source.ts:23`.
  - **The homepage build, capped, with and without the published source** (`homepage/public/source`,
    2,893 files, 70 MB), a clean `.next` each time:

    | Build | Time | Max RSS |
    |---|---|---|
    | with the source | 14.55 s, 15.15 s | 807,332 KB, 801,948 KB |
    | without it | 15.11 s | 797,744 KB |
    | `main`'s join, with the source | 13.99 s | 837,144 KB |

    No measurable difference: the opt-out is by rule, not a measured fix today.
- **Left:** F2 (an animation of `style` or `class` is not held to the style allowlist; the key's
  clip contains it). F6 (the form round-trip trims a stored SVG, so a hand-edited icon with
  surrounding whitespace is re-checked on a form save). F7 (at 390 px the chart's thinnest upper
  strata; see the gap below, which leaves a thin band its colour).

#### The chart's surface gap (ruling 1)

- **The growth chart** (`ArchiveGrowthChart.tsx`, `.growth-gap` in `homepage/app/globals.css`):
  - a 2 px gap along each band's upper edge, in `var(--background)` (the chart sits on the page
    ground, as the e2e checks); non-scaling, round joins; `Canvas` in forced colours;
  - drawn centred after every fill, so each neighbour gives 1 px;
  - no gap at the stack's top (the surface is already there);
  - one set of gaps per plot height (200, 260, 300 px), each in a `<g>` shown by its class: a gap
    is drawn only where both bands are at least 3 px tall at that height, so a band that gives
    one keeps at least 1 px of its colour and a thinner band touches its neighbour instead
    (after the review, measured at right angles to the edge: M1 in "Review" below);
  - a run under three months is dropped.
- **Measured on the published summary** (the first, vertical rule; after M1 see "Review"):

  | Plot height | Smallest band that gives a gap keeps | Smallest band with no gap | Month boundaries parted / touching |
  |---|---|---|---|
  | 200 px (a phone, 390 px) | 1.00 px | 0.10 px | 182 / 277 |
  | 300 px (1280 px) | 1.00 px | 0.15 px | 302 / 157 |

  At 390 px the 2016–2020 bands are 0.1–5.5 px tall.
- **The shared charts** (`common/components/charts/surfaceGap.ts`, `--chart-gap` in
  `tokens.css`: the chart surface, `Canvas` in forced colours):
  - stacked bars get a 2 px stroke per segment;
  - stacked areas got a 4 px stroke along the band's top edge, half under the band above, so
    2 px showed — withdrawn after the review (H1): they keep their series-coloured edge;
  - single-series charts, lines and side-by-side bars are unchanged.
  - This covers `ChartView` (the export's charts) and `CrossSiteChart` (the homepage's `/stats`
    customize view).
- **Screenshots:** `~/reports/release-14/shots/chart/gap2-{390,1280}-{light,sepia,dark}{,-thin}.png`
  (taken before T1 removed Sepia).
- **Departures for the operator to rule on, not changed:**
  - `ChartView` and `CrossSiteChart` fill stacked areas translucent (0.2 and 0.25); the standard
    is opaque fills parted by the gap.
  - `CrossSiteChart`, `LeaderboardChart` and `MomentumChart` draw dashed gridlines
    (`strokeDasharray="3 3"`); the standard is solid hairlines.
  - `ChartCard`'s surface is `--card`, equal to `--chart-surface` on both bases today.

### Slice T1, as shipped — two grounds (2026-09-28)

- **Two grounds.**
  - `THEME_BASES` is System, Light, Dark.
  - `nextBase` walks `THEME_BASES` (system → light → dark → system).
  - `ResolvedBase` is light | dark; `isThemeBase` takes the three.
- **Deleted:**
  - the Sepia block in `common/styles/tokens.css` and its `--base-sepia` flag;
  - `onSepia` in every accent, `BASE_GROUNDS.sepia`, `ACCENT_INK.sepia`, the fitted
    `--accent-custom-sepia`;
  - the BookOpen icon;
  - Sepia from every spec, fixture and comment;
  - `ThemeMenu.tsx` (nothing renders it);
  - `accentOptions` and `isThemeAccent`;
  - the `pinAccent` option (it is the only behaviour now).
- **Kept:**
  - `ACCENT_KEY`, documented as a pick nothing reads or deletes.
  - `ThemeRadios`, base-only in this commit; H2 deleted it with its caller.
- **Migration, no flash:**
  - A stored `ytdlp-tb:base` of the retired value (`RETIRED_BASE` in `themeConfig.ts`, the one
    place its name is spelled) is `light` in the pre-paint script and in `ThemeProvider`, and is
    rewritten to `light` once.
  - The legacy `archive` + `light` pair maps to `light`.
  - Any other unknown value falls back to the app's default, as before.
  - Unit tests cover the whole matrix (the retired value, `archive` + `light`, garbage, a storage
    that will not take the write).
  - Each app has an e2e for a stored retired base: an init-script MutationObserver sees no ground
    but the server's and Light, `data-theme-ready` is set, and storage reads `light`.
- **The accent is the site's:**
  - `ThemeScript` and `ThemeProvider` never read `ytdlp-tb:accent` and never remove it.
  - The export's layout and header, and the editor's header, have no accent picker; the
    slide-out menu's accent list is gone.
  - The site form's accent control is config and is unchanged; its hint no longer says a reader
    can pick another.
  - `theme-accent.spec.ts` keeps the site's own accent (fitted per base, no picker), a stored
    accent ignored with no flash and left in place, and the hint in the site's accent on each
    base. The four-bases menu, "pick Violet" and "pick the site's colour again" went with the
    picker.
- **Docs:** `operate.md` (no theme menu; a light or dark ground with the header's toggle),
  `siteSchema`'s accent text (`SITE.md` regenerated), the site form's hint.
- **Counts** (`git grep -i sepia -- . ':!plans/' ':!*CHANGELOG.md'`): 47 files, 177 lines at the
  branch's start (with HP merged). After T1: one line, `RETIRED_BASE = "sepia"`, which the
  migration needs in order to recognise a stored value. The changelogs' `[Unreleased]` sections
  name it nowhere; released entries keep 9 lines.

### Slice H1/H2, as shipped — the export and editor headers (2026-09-28)

- **The export header** (every site and the hub):
  - From `lg` (1024 px): brand · nav · **Archilyzer** · the group.
  - Below `lg`: brand · the group · the menu trigger.
  - The group is the homepage's: `SocialLinks` (header, at most four: `featured`, else the last
    four) in `SocialScroll`, then `ThemeToggle variant="bare"`, 36 px keys (44 px under a coarse
    pointer) touching.
  - The footer's row is `SocialLinks` (footer) too: all the links, the same keys, the text colour
    on hover where it was the accent.
- **The Archilyzer link** replaces `SiblingSwitcher` (deleted):
  - visible text "Archilyzer", accessible name "Archilyzer — official instances";
  - `INSTANCES_URL`, same tab;
  - not on the hub.
  - The homepage's Official Instances section has `id="instances"`, with `scroll-mt` equal to the
    sticky header's height. The planned H3 is folded in here, with a homepage e2e for
    `/#instances`.
- **The header's Hub link** is gone from the bar and the menu; `hubUrl` and `resolveHubUrl` still
  parse. **Changelog** is in the footer after Use with AI, and not in the bar or the menu.
- **The slide-out menu** holds only the nav, plus the Archilyzer link on a site. It is kept:
  five-plus nav links do not fit a phone's bar. `ThemeRadios` is deleted.
- **The nav's breakpoint moved from `md` to `lg`:**
  - At 768 px the nav, the Archilyzer link and a long title with four icons did not fit, so the
    wordmark hid and an icon scrolled while the nav was still inline.
  - With the nav in the menu below `lg`, the icons never scroll at 768.
- **Narrow widths, for any title** — the approach, and why (superseded in part by ruling 5:
  below 520 px the header now shows only the marked links, and the wrap and the scroll box are
  last resorts; see "The narrow header, as ruled"):
  - The wordmark's text sits in a one-line box (`h-7`, `overflow-hidden`, `flex-wrap`) behind a
    zero-width strut. The text is one flex item: when it does not fit beside the strut, it wraps
    to the second line, which is clipped.
  - So the text drops exactly when it does not fit, measured by the browser in the site's own
    font at the reader's text size. There is no per-site threshold and no build-time estimate.
    Titles differ in length, and a threshold written down would be wrong for some title or some
    text size.
  - The text stays in the DOM, so the link's name stays the title.
  - Below `lg` the brand link takes the bar's free space (basis 0, grow 1, minimum the mark), so
    the group gives way only once the link is the mark alone. The row's box then scrolls, end
    first, as on the homepage.
- **Measured** on a dev server (synthetic titles, the viewport width):

  | Title | Links | Full wordmark from (mouse / touch) | Row scrolls below (mouse / touch) |
  |---|---|---|---|
  | "Shortlyzer" | 1 | 321 / 337 | never ≥ 240 |
  | | 2 | 357 / 381 | never / 251 |
  | | 3 | 393 / 425 | 263 / 295 |
  | | 4 | 429 / 469 | 299 / 339 |
  | "Longestfixturealyzer" | 1 | 449 / 465 | never ≥ 240 |
  | | 2 | 485 / 509 | never / 251 |
  | | 3 | 521 / 553 | 263 / 295 |
  | | 4 | 557 / 597 | 299 / 339 |

  The page never scrolls sideways from 240 to 1023 px. The row scrolls at 320 px only with four
  links under touch (by 19 px): the export bar also carries the menu trigger, which the
  homepage's does not.
- **The editor's header:**
  - `ThemeToggle` (its default variant, the editor's existing chrome) cycles System, Light and
    Dark.
  - `ThemeMenu` is gone.
  - Nothing else is restyled.
  - There is no social row.
- **Tests:**
  - `header.spec.ts`, new:
    - a short and a long title × 1–4 links × 280–430 px × mouse and touch, checking that the page
      and the header never overflow, every link shows, the text drops before the row scrolls, and
      the mark, the toggle and the link's name hold;
    - a short title in full at 390 px, a long one giving way;
    - key sizes (36 and 44 px) and touching boxes;
    - the focus ring;
    - the Archilyzer link's href, name and text; no sites button, hub link or "Choose theme";
    - Changelog in the footer after Use with AI and not in the header;
    - six links and two featured;
    - a hostile icon in `socialLinks` as its label, running nothing and requesting no other
      origin.
  - The fixture site is rewritten per test and restored. It carries its original in
    `_e2ePristine`, which `playwright.config.ts` restores if a run dies.
  - `responsive.spec` covers the menu's contents; `site-branding` scopes its footer link;
    `official-instances` (hub) checks the hub's header has no Archilyzer link, sites menu or theme
    menu.
- **Screenshots** (`~/reports/release-14/shots/h1/`, synthetic icons, six configured so the header
  shows four):
  - `{short,long,hub}-header-{320,360,390,768,1280}-{light,dark}.png`;
  - `{short,long,hub}-footer-{…}.png`;
  - `editor-header-{1280,390}-{light,dark}.png`.

#### Gates (at `0f358ee7`; logs `$T/t-g-*.log`)

- **tsc** was clean before every commit, and at the tip (49 s).
- **Unit:**

  | Suite | Result |
  |---|---|
  | common | 2,204/2,204 |
  | editor unit | 87/87 |
  | homepage unit | 8/8 |
  | `test:scripts` | 191 passed, 1 skipped (192) |
  | mcp | 271/271 |

- **Docs:** `settings example --check`, `docs files --check` and `docs env --check` all exit **0**.
- **Builds**, each capped at 5 GB with no swap, one at a time, from a clean `.next` (time, max RSS):

  | Build | Time | Max RSS |
  |---|---|---|
  | homepage (fixture icons) | 17 s | 796 MB |
  | export, site (the worktree's default) | 25 s | 982 MB |
  | export, the fixture site | 29 s | 990 MB |
  | export, hub | 26 s | 1,004 MB |
  | editor, with the corpus visible | 44 s | 1,645 MB |
  | umtool, with the corpus visible | 19 s | 784 MB |

  For the editor and umtool builds the primary's `transcripts/` was linked in, and the worktree's
  own set aside. Both were put back after the builds; nothing ran through the link.
- **e2e**, each detached and queued:

  | Suite | Passed | Failed | Time |
  |---|---|---|---|
  | homepage, full | 77 | 0 | 1.8 min |
  | export, full | 220 | 0 | 8.6 min |
  | hub, full | 34 | 0 | 1.1 min |
  | editor: `theme`, `sites-crud`, `settings`, `export-search` (the specs on the theme controls, the site form's accent hint, the social-links field, and the export's header and footer as the editor's export server shows them) | 51 | 0 | 1.7 min |

- **Along the way:**
  - the chart commit: tsc; homepage unit 8/8; a capped homepage build 16 s; homepage e2e 75/75
    (2.2 min); export `charts.spec` 9/9;
  - T1: homepage theme/toggle/brand/instance specs 21, export theme/branding specs 30, hub 8,
    editor 20;
  - H1/H2: export header/responsive/branding specs 43 (one assertion order fixed), hub 9,
    homepage `marketing` 11.
- **Numbers tool:** none.

#### Found and left

- **The export footer's `-mx` inset** keeps the shared row's value; on a wide screen the last
  key's box reaches 8 px into the container's padding.
- **At 320 px with four links under touch** the export header's row scrolls by 19 px. This is the
  ruled last resort; the menu trigger is what the homepage does not have.
- **`resolveHubUrl`** has no caller in the export app now. It still parses, as ruled.
- **Playwright's own Firefox and WebKit are not installed** at this Playwright's revisions; F1 was
  checked by hand in the installed Firefox 146.
- **The chart departures** above (translucent stacked areas, dashed gridlines) are for the
  operator.

#### Decisions the operator could overturn

| What I assumed | The alternative |
|---|---|
| The export's inline nav and the Archilyzer link start at `lg` (1024 px), with the menu below | keep `md` and let the wordmark hide and an icon scroll at 768–1023 px |
| The Archilyzer link is in the slide-out menu below `lg` | leave it out of the menu, so a phone reaches it only by the footer's "Built with Archilyzer" |
| The link is styled as the old Changelog link (small, muted), with the text "Archilyzer" | mono uppercase like the old Hub link, or the nav's style |
| The wordmark's text hides by a pure-CSS wrap, exact for any title | a per-site threshold computed at build time from the title's estimated width |
| The export footer's icons take the shared keys (36 px, hover to the text colour) | keep the footer's 20 px accent-hover icons |
| The chart's gap is skipped where either band is under 3 px at the drawn height, and runs under three months are dropped | a gap on every boundary, erasing the thinnest bands |
| Stacked areas in the shared charts get the surface gap over their translucent fills | leave their series-coloured top lines until the fill opacity is ruled on |
| `RETIRED_BASE = "sepia"` stays in `themeConfig.ts` for the migration | spell it indirectly so the grep reads zero |
| The export footer's dot shows only after a downloads link | leave the dot unconditional, as before |

#### Review (verdict SHIP AFTER FIXES; `$T/t-review.md`)

| Finding | Fix |
|---|---|
| H1: the stacked-area surface gap erased small values and cut peaks (every site's multi-series area charts, `/stats`) | `92dc4084`: stacked areas keep `main`'s series-coloured edge (ChartView 1 px, CrossSiteChart 1.5 px); `BAR_GAP` stays on stacked bars. `charts.spec` reads the chart back from a screenshot (`common/testing/chartPixels.ts`): at 390 and 1280 px the stack's topmost painted row at each month is within 1 px of the value scale's y for the true total, and each series shows pixels of its own fill; with the old gap the tops land 2–4 px low |
| M1: the growth chart's thin-band rule was vertical, so on steep edges the gap covered whole bands | `f455d7da`: `homepage/app/lib/growthGaps.ts` (pure): thickness at right angles to the edge (min vertical height × cos θ) at the narrowest plot of each height (240, 592, 976 px); a gap only where both bands keep ≥ 1 px after every gap on them. `growthGaps.test.ts` checks every drawn segment independently: 0 covered band-segments on the fixture, a slivers-and-spikes stand-in and a one-month spike (two of the four fail under the old rule). `growth-chart.spec`: only the drawn height's set shows; the peak's painted top is within 1 px of the true total; every site with data shows its colour |
| M2: four texts still promised the Hub link | `706f62bf`: Settings' Family hub URL and a site's Hub URL hints, `homepageUrl`'s and `hubUrl`'s docs: the value is published as `hubUrl` in the site's `/site.json` and `/corpus.json` so the hub can tell member sites; no page links to it. `SETTINGS.md`, `SITE.md` regenerated; labels and field names unchanged |
| L1: from 1024 to 1046 px a very long title dropped its text and scrolled the icons at once | `2a5a81c0`: `lg:shrink-[999]` on the brand link; `header.spec` sweeps 1024–1050 px under touch with a 27-character title (scrolls without the fix) |
| L2: a title could show in the fallback face and vanish when Archivo loaded | `2a5a81c0`: the wordmark's width is reserved from Archivo's metrics at its two instances (`common/lib/wordmarkMetrics.ts`, generated by `common/bin/gen-wordmark-metrics.py` from the vendored font; `lib/wordmarkWidth.ts`, +3 %: the served font renders 0.2–1.7 % wider than the file's advances). `header.spec` blocks the font file and compares the title's visibility at 360, 375, 390, 412 and 430 px: the same before and after Archivo loads (the short title differs without the reservation) |
| L3: the provider and the script disagree when storage reads but refuses writes | left: reachable only with a stored retired value that can no longer be rewritten |
| L4: `common/components/ui/` was not untouched | nothing to change: `alert.tsx`, `badge.tsx` and `sonner.tsx` have comment-only edits naming the bases; the palette, `dialog.tsx` and `command.tsx` are untouched |
| L5: stale text | `689a835f`: `export/globals.css`, `export/manifest.ts`, `brand.test.ts`; the export changelog's T1 bullet no longer names a Base list; the homepage changelog's e2e bullet |
| L6: the footer's social row did not wrap | `2a5a81c0`: `flex-wrap` on the footer placement; eight links under touch at 320 px, export and homepage specs |
| L7: the guard's blind spots (`fs.promises.*`, `accessSync`/`open`/`readlink`, a relative literal to fs, a cwd default parameter, a class method as the source; false positives on a string naming `process.cwd()`) | left: none of these shapes is in the covered code today |
| L8: the rollout order was implied | this commit: "## Rollout" below |
| L9: 200 % text | `2a5a81c0`: the header itself never overflows at 200 % text (tested at 320 and 360 px, both pointers). The 43 px the review measured at 320 px are the search page's own controls (a label and the view toggle), not the header. At 320 px under touch the mark, the 88 px toggle and the 72 px menu button leave the social row's box no room: the phone header is the operator's to rule on |
| L10: named-accent coverage | `689a835f`: `theme-accent.spec` rewrites the fixture to a named accent (violet): its value on Light and on Dark, a stored pick not applied and left in place |

**The growth chart's gaps after M1**, on the published summary:

| Plot height | Gap segments | Runs | Covered band-segments |
|---|---|---|---|
| 200 px | 25 | 5 | 0 |
| 260 px | 64 | 11 | 0 |
| 300 px | 96 | 14 | 0 |

**The gaps now read as dashes.** On a phone the chart has two short dashes, both along
Jeralyzer's top. At 1280 px they are stretches along Jeralyzer's and Anilyzer's tops, with
short pieces elsewhere. The rest of the bands touch. This is the underlying issue the operator
has been told about: four sliver bands beside two large ones. Folding the small sites into
"Other", or small multiples, is the operator's call and is not made here.
Screenshots: `~/reports/release-14/shots/chart2/growth-{390,1280}-{light,dark}{,-2019-2022}.png`.

**Screenshots:**
- `chart2/site-stacked-{area,bar}-{390,1280}-{light,dark}.png`: a site's charts. The fixture has
  one video a month, so its stacked bars have no touching segments.
- `chart2/stats-stacked-{area,bar}-{390,1280}-{light,dark}.png`: `/stats` on the published
  summary. The area has its coloured edges back, and the bars show the 2 px gaps between
  segments.

**The nine decisions, as ruled:**
1. The nav and the Archilyzer link are inline from `lg` → keep, with L1.
2. The Archilyzer link is in the slide-out menu below `lg` → keep.
3. The link is small and muted, with the text "Archilyzer" → keep.
4. The pure-CSS wrap → keep, with L1 and L2.
5. The export footer takes the shared 36 px keys → keep, with L6.
6. The growth-chart gap is skipped under 3 px, and runs under three months are dropped → keep
   the rule, measured at right angles (M1).
7. The surface gap over translucent stacked areas → do not keep (H1).
8. `RETIRED_BASE = "sepia"` stays → keep.
9. The footer's dot only after a downloads link → keep.

| sha | what |
|---|---|
| `92dc4084` | H1 |
| `f455d7da` | M1 |
| `706f62bf` | M2 |
| `2a5a81c0` | L1, L2, L6, L9 |
| `689a835f` | L5, L10 |
| `447ded9a` | the narrow-header ruling (below) |
| _this_ | `plans:` the Review, the ruling's record, the Rollout, the plan, STATE |

**Gates**, run once after the review fixes and the ruling (at `447ded9a`; logs `$T/t-g3-*.log`):

- **tsc** was clean before every commit, and at the tip (37 s).
- **Unit:**

  | Suite | Result |
  |---|---|
  | common | 2,210/2,210 |
  | editor unit | 87/87 |
  | homepage unit | 12/12 |
  | `test:scripts` | 191 passed, 1 skipped |

- **Docs:** the three `--check` commands exit **0**.
- **Builds**, each capped at 5 GB with no swap, one at a time:

  | Build | Time | Max RSS |
  |---|---|---|
  | homepage | 16 s | 800 MB |
  | export site | 25 s | 995 MB |
  | export hub | 23 s | 1,021 MB |
  | editor | 43 s | 1,650 MB |

- **e2e**, each detached and queued:

  | Suite | Passed | Failed | Time |
  |---|---|---|---|
  | homepage, full | 90 | 0 | 2.1 min |
  | export, full | 235 | 0 | 10.9 min |
  | hub, full | 35 | 0 | 1.3 min |
  | editor (`settings`, `sites-crud`, `theme`) | 32 | 0 | 1.3 min |

- **An earlier gate run** at `689a835f` (the review fixes alone) was stopped when the ruling
  arrived, to run the gates once for both. Before it stopped, it had passed: tsc; common
  2,208/2,208; editor unit 87/87; homepage unit 12/12; `test:scripts` 191 + 1 skipped; the docs
  checks; all four builds; and the homepage e2e, 83/83.

#### The narrow header, as ruled (2026-09-28)

**The ruling** (operator): on narrow screens a site's header keeps the site's NAME and shows only
the link(s) marked for the header; the other social links are in the footer.

**`featured`, new meaning** (`common/lib/socialLinks.ts` `headerSocialLinks(links, width)`):
- **Wide:** every link, up to four. With more configured, the `featured` ones are kept first, then
  the last of the rest, shown in configured order.
- **Narrow:** only the `featured` ones, up to four (the last four when more are marked). With none
  marked, the narrow header shows no social link: the name wins, and the footer has them all.
- **The footer** always shows every link.
- The key is still `featured`, stored only when true. The editor's checkbox is **Keep in header
  on small screens**, with the hint "On small screens the header shows only these; the rest stay
  in the footer." `SETTINGS.md` and `SITE.md` are regenerated.

**Layout** (the homepage, every site and the hub, one behaviour):
- Both rows are rendered. CSS shows the narrow one below **32.5rem** and the wide one from
  32.5rem: 520 px at the default text size, 1040 px at a 32 px browser text size (`11a33f7a`, N5).
  The other is `display: none`, so exactly one is focusable and in the accessibility tree.
- Each copy scopes its icons' ids.
- The narrow bar is mark + name · the marked link(s) · toggle, then the menu button on the export.
- The wordmark's text drop (the export's reserved-width wrap, the homepage's container
  thresholds, now keyed by the narrow row's count) and the scroll box stay as last resorts.

**Why 520 px:** every link (up to four, 44 px touch keys), the toggle, the menu button and the
longest real title (Rekietalyzer, Hasanalyzer) fit the export's bar from about 500 px wide under
touch. At 519 px the narrow row shows and at 520 px the wide one, and the name shows at both
(measured on Rekietalyzer and Hasanalyzer, both pointers, 519–1280 px). The homepage's bar needs
about 424 px.

**The gap ruling (2026-09-29):** below the switch the export bar's two gaps (name → row, toggle →
menu) are **8 px**, not 12 (`1a2e3342`, N1). Type, the mark–name gap and the reserved width's
3 % margin are unchanged. It is one header, so every site and the hub. The homepage's bar has one
gap and keeps 12 px.

**The smallest viewport width at which the full name shows** (a dev server at `11a33f7a`, four
links configured, re-measured after the gap ruling; before it, each site's and the hub's widths
were 8 px more):

| Title | One marked link: mouse / touch | None marked: mouse / touch |
|---|---|---|
| Jeralyzer | 302 / 318 | 266 / 274 |
| Anilyzer | 288 / 304 | 252 / 260 |
| Bonnellyzer | 336 / 352 | 300 / 308 |
| Hasanalyzer | 343 / 359 | 307 / 315 |
| Rekietalyzer | 344 / **360** | 308 / 316 |
| Jasolyzer | 308 / 324 | 272 / 280 |
| Archilyzer (homepage) | 276 / 292 | ≤ 240 / 248 |
| Archilyzer (hub) | 313 / 329 | 277 / 285 |

**The target** (every title in full at 360 px, one marked link, touch) is **met by every title**.
Rekietalyzer meets it at exactly 360 px.

**The wordmark's reservation is a minimum width** (`4ea1c495`, N2). The header sets
`wordmarkWidthEm` as `min-width`. Where the text renders wider than the reservation, its box grows
with it and the text drops sooner; its last letter is never clipped. The re-review measured 0.6 to
0.8 % spare in Chromium. The table above is unchanged by it.

**`featured` keeps its configured place** (N4, ruled as built). With more than four links, a
marked link is guaranteed a place in the wide row, and the row keeps configured order: `[A★, B,
C, D★, E, F]` shows `A D E F`.

**Tests:**
- `socialLinks.test.ts` covers both widths: none marked, one, several, more than four with and
  without marks, and order kept.
- The homepage (`social.spec`), a site (`header.spec`: a short title, Bonnellyzer, Hasanalyzer and
  Rekietalyzer) and the hub (`official-instances.spec`) are checked at 360 and 390 px, both
  pointers:
  - the full name shows;
  - only the marked link is in the header, in the accessibility tree too;
  - with none marked, the header has no row and the name shows;
  - from 520 px every link shows;
  - the footer has every link at every width;
  - nothing scrolls the page sideways from 280 to 1400 px;
  - focus goes brand → marked link → toggle, then menu on a site.
- `header.spec` checks **all six real titles at 360 px under touch with one marked link**: the
  full name, the link, no scrolling row, no overflow. Hasanalyzer and Rekietalyzer fail it with
  12 px gaps.
- `header.spec` forces the wordmark 8 % wider (letter-spacing) and sweeps 519 to 280 px:
  wherever the name shows, it does not overflow its box or the clip box. With a fixed width it
  is clipped from 519 to 344 px.
- **At 200 % text**, both pointers:
  - **The browser's text size** (32 px, set through CDP `Page.setFontSizes`; Chromium):
    - the narrow row shows until 1040 px, and the wide one from there;
    - the row never scrolls while the name shows, and the header never overflows;
    - the name shows with one marked link from 720 px on the homepage and from 767 px on
      Rekietalyzer.
  - **The page's root at 200 %** (which moves the page's rem but not a media query's):
    - on the homepage from 520 to 767 px, the wide row never scrolls while the name shows;
    - on a site the switch stays at 520 px and the row never scrolls while the name shows.
- The e2e fixtures mark their last link.

**Screenshots:** `~/reports/release-14/shots/h2/`, four synthetic icons, the last marked:
- `{jeralyzer,anilyzer,bonnellyzer,hasanalyzer,rekietalyzer,jasolyzer,homepage}-{320,360,390}-{light,dark}.png`
  (a mouse) and `…-{320,360,390}-touch-{light,dark}.png` (touch), re-taken at `11a33f7a`;
- `…-{768,1280}-{light,dark}.png` from `447ded9a` (the header is unchanged at those widths).

At 360 px every title shows in full under touch. At 320 px Bonnellyzer, Hasanalyzer and
Rekietalyzer drop their text under both pointers, and Jasolyzer does under touch.

#### Re-review (verdict SHIP; `$T/h-review2.md`)

| Finding | Fix |
|---|---|
| N1: at 360 px under touch, with one marked link, Hasanalyzer and Rekietalyzer lost the name | `1a2e3342`: the gap ruling (above). All six titles, the homepage and the hub re-measured; the e2e checks all six |
| N2: the reserved wordmark width had 0.6–0.8 % spare in Chromium | `4ea1c495`: the reservation is a `min-width` (above) |
| N3: the Rollout had no step for marking the link, and a live check expected every icon at 390 px | this commit: a precondition before any build, and the 390 px check reads "the marked link" (## Rollout) |
| N4: "kept first" is priority, not position | ruled as built: a marked link keeps its configured place (above) |
| N5: on the homepage at 150–200 % text, from 520 to 767 px, the wide row scrolled while the name showed | `11a33f7a`: the switch is `32.5rem` in every header; the homepage's wordmark fit is keyed by the narrow row's count below the switch and by the wide row's from it |

| sha | what |
|---|---|
| `1a2e3342` | `export:` the bar's two gaps are 8 px below the switch; `header.spec` all six real titles; the hub's name test under both pointers (N1) |
| `4ea1c495` | `export, common:` the wordmark's reservation is a `min-width`; the wider-text e2e (N2) |
| `11a33f7a` | `homepage, export:` the switch in rem; the homepage's fit by the wide row's count from the switch; the 200 % text e2e (N5) |
| _this_ | `plans:` the ruling's record re-measured, the re-review, the Rollout's marked-link precondition (N3), STATE, the plan; the changelogs |

**Gates** at `11a33f7a` (logs `$T/t-g4*.log`):
- **tsc** was clean before every commit, and at the tip (32 s).
- **Unit:** common 2,210/2,210; homepage 12/12.
- **e2e**, the header specs only, detached and queued:

  | Suite | Specs | Passed | Failed |
  |---|---|---|---|
  | homepage | `social`, `toggle`, `svg-vectors`, `brand` | 50 | 0 |
  | export | `header`, `responsive`, `site-branding` | 50 | 0 |
  | hub | `official-instances` | 8 | 0 |

- **Builds**, capped at 5 GB with no swap, one at a time: the homepage (13 s, 824 MB) and the
  export site (21 s, 1,029 MB). Both builds' CSS carry the `32.5rem` switch, and the homepage's
  carries both fit sets.

**Found and left:** with the page's root at 150–200 % (not a browser setting), from 768 px the
homepage's nav joins the bar at `md`, whose media query does not follow a page-set root size.
There the bar overflows and the wide row scrolls while the name shows: measured from 779 to
1100 px at 150 % and from 779 to 1300 px at 200 %. With the browser's own text size `md` moves too
and nothing overflows. The `md` layout is slice HP's and unchanged here.

### Slice HS, as shipped — hidden sites: a site that builds and deploys, and is listed nowhere (2026-09-29)

Branch `r14/hidden-sites` off `main` `99d4d76a`, worktree `~/Projects/homepage-social-visible`
(block #3: editor test 3311, export 3310, homepage e2e 3340, hub e2e 3341), one Opus implementer.
Scratch files `hs-*` in the job's `tmp`. The rulings (2026-09-29, not re-opened):
1. A hidden site is absent from the homepage (cards, chart, `/stats`), the hub (members, federated
   search, `corpus.json`, `llms.txt`), every other site's related-sites footer, and the published
   id lists (`channel-sites.json`, the pooled `stats/`).
2. Its hours and transcripts are in no public total.
3. The editor has one checkbox in each site's settings, and no homepage settings page.
4. The hidden site itself builds and deploys exactly as before.

| sha | what |
|---|---|
| `122b8879` | `common:` `site.json` `listed` (schema, docs, parser, writer; `SITE.md`); `isListedSite` and `channelsOnlyOnUnlistedSites`; the footer drops an unlisted sibling |
| `77aa1456` | `common:` the homepage summary (v6), `channel-sites.json` and the whole-pool stats leave out an unlisted site and the channels only it exposes |
| `cbd12a8b` | `common:` `hub-sites.json` (and through it the hub's `corpus.json` and `llms.txt`) lists listed sites only |
| `5836d6ec` | `sites:` the checkbox, and `saveSiteAction` carries the key; `sites-crud` e2e |
| `64ae9b47` | `homepage:` the e2e fixture's seventh, unlisted site; `unlisted-site.spec.ts` |
| `30b3f486` | `plans:` this record; the three changelogs |
| `e3ee2eb1` | `homepage:` the fixture's second list unlists every site in it; the spec proves the site is in the input (review L1, L2) |
| `65675373` | `common:` a shared channel stays the listed site's when the unlisted id sorts first (review L3) |
| `20ff9ecf` | `plans:` FACTS' naming hazards (L4); this release's slices table, Order and Rollout carry HS (L6) |
| `7ea35dfb` | merge of `main` `ccf90892` (release 15 IG); `buildStats.ts` and `editor/CHANGELOG.md` merged clean, the HS bullet under `[Unreleased]` |
| _this_ | `plans:` the commit table, the review and the post-merge gates |

The six commits up to `30b3f486` were rewritten after the review for the release's commit trailer
(`git filter-branch --msg-filter`, trees unchanged); their first shas were `5918ba27`, `10378f7f`,
`1491a7ab`, `ca22f0e6`, `e12515c2`, `4eeb80ee`.

- **The key:** `listed?: boolean` in `site.json`, after `siteUrl`, in the type, `SITE_FIELD_DOCS`,
  `siteFieldsSchema` and `siteToDisk`. Absent or anything but `false` reads `true`; only `false` is
  written (the `archives` idiom). `SITE.md` regenerated.
- **One predicate.** `isListedSite(site)` (`site.listed !== false`) and
  `channelsOnlyOnUnlistedSites(sites)` — the channels at least one site exposes and no listed site
  does. Both live beside the key in `common/lib/siteSchema.ts` and are exported from `lib/site`
  (its `export *`), like `isValidSiteId`: the summary builder is pure and imports them without
  `lib/site`'s file I/O. Every filter below calls them.
- **What each public output does now:**

  | Output | Code | An unlisted site | A channel only unlisted sites expose |
  |---|---|---|---|
  | Homepage summary: `sites`, `official`, `monthly`, `series`, `recent`, `channels` | `buildHomepageSummary`, the public-site filter | absent | absent |
  | Homepage summary: `totals`, `availability` | the same, over the in-scope stats | — | not counted |
  | `channel-sites.json` | `channelSitesOf` (`controller/poolSummary.ts`) | absent from every list | absent |
  | The homepage's `stats/` (whole-pool bundle) | `buildStats`, the whole-pool block | — | its records and its manifest entry absent |
  | `hub-summary.json` | `toHubSummary` of the same summary | absent | not counted |
  | `hub-sites.json`, the hub's `corpus.json`, `llms.txt` | `compose-hub.ts`, the built-in pool | absent | — |
  | Every other site's footer | `resolveRelatedSites` | not linked, even from a featured group | — |

  - A channel a listed site also exposes is credited to the listed site: `primarySiteOf` runs over
    listed public sites only.
  - The summary's version is 6. No field was added or removed; the number marks the scope change.
- **What stays, and why:**
  - The unlisted site's own build and deploy: `buildAll` and the deploy paths
    (`common/publish/build.ts`), its per-site stats and summaries (`buildStats`, `buildIndex`), its
    archives and chart templates, its own `/site.json`, `corpus.json`, `llms.txt`, sitemap. Its own
    footer still lists its listed siblings.
  - The editor's pages (`/sites`, the channel page's memberships, the nav's site switcher, the
    priority focus), which list every site.
  - `siteChannelIndex`, `renameChannel`, `migrateToSites`, `listSiteIds` for the CLI's
    site argument, the export's default site: none is a public list.
  - The per-site staging in `buildIndex.ts` (another slice's file, and per site).
- **Editor.** The site form has **List on the Archilyzer homepage and hub** after Public URL, on
  by default, with a hint. `saveSiteAction` rebuilds the `Site` from the form; it now carries
  `listed: false`, so a save of any other field keeps it. The other writers (`writeSite` from the
  channel page, `renameChannel`, `migrateToSites`) spread the stored site or create a new one.
- **Tests:**
  - `siteSchema.test.ts`: absent and `true` read listed, only `false` unlists and only `false` is
    written, through `writeSite`/`getSite`; the "everything" round-trip fixture carries
    `listed: false`; `channelsOnlyOnUnlistedSites` keeps a shared channel with the listed site; the
    footer drops an unlisted sibling named in a featured group, and an unlisted site's footer
    still lists the rest.
  - `homepageSummary.test.ts`: an unlisted site with its own channel and one shared with a listed
    site gives a summary deep-equal to the one without it (every array, every total), and its id,
    title and channel appear nowhere in the JSON; an unlisted site with no `siteUrl`, and one whose
    channels are all shared, change nothing either; `listed: true` equals no key.
  - `buildStats.test.ts` (k): the whole-pool bundle leaves out the unlisted-only channel (records,
    manifest, count, log line) and keeps a pool-only one; the unlisted site's own bundle keeps all
    three of its videos.
  - `poolSummary.test.ts` (new): `channelSitesOf` names listed sites only.
  - `compose-hub.test.ts`: `hub-sites.json`, `corpus.json` and `llms.txt` name the listed site and
    not the unlisted one.
  - Editor `sites-crud`: a `listed: false` file opens unticked; a save that changes only the title
    keeps `false`; ticking removes the key; unticking writes it again.
  - Homepage `unlisted-site.spec.ts`: the builder's input holds the unlisted site (`listed:
    false`, its own records, the shared channel), and listed it would add a card and its records;
    the summary the dev server reads equals the one built without it; `/` and `/stats/` (served
    HTML and DOM) name all six listed sites and not the unlisted one.
  - `homepageSummary.test.ts`, after the review: an unlisted id that sorts BEFORE the listed one
    (`aaa-hidden` < `beta`) still leaves the shared channel with the listed site.
- **The homepage e2e fixture** (`homepage/e2e/fixture-summary.ts`), for the slices that build on
  it:
  - `FIXTURE_SITES`: the six listed sites, unchanged.
  - `FIXTURE_UNLISTED_SITE`: `fixture-unlisted`, "Fixture Unlisted", two channels of its own
    (`fixture-unlisted-ch1/2`, six a day, like the rest); it also exposes `fixture-one-ch1`.
  - `buildFixtureInputs(fixtureSites = FIXTURE_SITES, unlistedSites = [FIXTURE_UNLISTED_SITE])`
    returns the builder's inputs (`stats`, `channelSites`, `sites`). The first list is listed;
    every site in the second is written with `listed: false`, whatever it carries (`FixtureSite`
    has no `listed` field), and shares the first listed site's first channel.
  - `buildFixtureSummary(…)` is the real builder over those inputs. The unlisted sites' records are
    generated last, after the megaspike, so every listed record is the same with them or without,
    and `buildFixtureSummary()` equals `buildFixtureSummary(FIXTURE_SITES, [])`. Passed in the
    FIRST list, the same site is listed: 7 sites and 39,599 transcripts instead of 6 and 37,199.

#### Proof: a hidden fixture site through the real builds

A throwaway corpus in the job's scratch dir (`$T/hs-proof/`, `make-corpus.mjs`): three channels
of three captioned videos each, and two sites with public URLs — `fixture-listed` (the listed
channel and the shared one) and `fixture-unlisted` (`listed: false`; the shared channel and one of
its own). Every path the builds write was pinned there (`TRANSCRIPTS_DIR`, `SETTINGS_FILE`,
`EXPORT_INDEX_DIR`, …) except the two apps' own `public/` and `out/`. The worktree's own gitignored
`homepage/public` data and `export/out` were set aside first and put back after; the
`export/public` links were dropped (never their targets) and re-seeded from the primary after. The
primary's `export/public` was untouched (no entry newer than the slice's start). Each build was
capped at 5 GB with no swap.

| Build | Result | Time | Max RSS |
|---|---|---|---|
| `archilyzer index` | 0 | 4 s | — |
| `archilyzer build homepage --no-source` | 0 | 20 s | 776 MB |
| `archilyzer build hub` | 0 | 34 s | 1,005 MB |
| `archilyzer build site fixture-unlisted --skip-archives` | 0 | 41 s | 962 MB |
| `archilyzer build site fixture-listed --skip-archives` | 0 | 84 s | 970 MB |

Counts (files holding the string / occurrences, `grep -rF`):

| Tree | `fixture-unlisted` | its title | its own channel's slug | its own channel's name | `fixture-listed` |
|---|---|---|---|---|---|
| `homepage/public` (summary, `channel-sites.json`, `stats/`) | 0 / 0 | 0 / 0 | 0 / 0 | 0 / 0 | 2 / 25 |
| `homepage/out` | 0 / 0 | 0 / 0 | 0 / 0 | 0 / 0 | 10 / 145 |
| `export/public` (hub compose) | 0 / 0 | 0 / 0 | 0 / 0 | 0 / 0 | 4 / 10 |
| `export/out` (hub) | 0 / 0 | 0 / 0 | 0 / 0 | 0 / 0 | 4 / 10 |
| `export/out` (the listed site) | 0 / 0 | 0 / 0 | 0 / 0 | — | 9 / 30 (its URL) |

- The compose lines: `compose-homepage: 2 channel(s) mapped across 1 listed site(s) (1 unlisted
  left out); summary covers 6 transcription(s) / 6 download(s) across 1 public site(s)` and
  `compose-hub: 1 built-in pool site(s) …; hub-summary.json covers 1 official instance(s)`.
- The summary: version 6; `totals` 6 transcripts, 6 downloads, 1 site, 2 channels, 6 hours;
  `official` the same; `availability.counted` 6. With the unlisted site counted they would have
  been 9 transcripts and 9 hours.
- The pooled `stats/` manifest: 6 records, channels `proof-listed-channel` and
  `proof-shared-channel`. `channel-sites.json` maps both to `fixture-listed` alone.
- **The unlisted site still builds:** its own stats bundle holds all six of its videos (both its
  channels); its build ships its own pages (its id in 17 files), and its footer links
  `https://fixture-listed.example`. The listed site's footer links nothing: its only sibling is
  unlisted.

#### Gates (at `64ae9b47`, the tree of the first `e12515c2`; logs `$T/hs-*.log`)

- **tsc** was clean before every commit (69 s, 33 s, 44 s — the last over the tip's code).
- **Unit:**

  | Suite | Result |
  |---|---|
  | common | 2,218/2,218 (8 new) |
  | editor unit | 87/87 |
  | homepage unit | 12/12 |
  | `test:scripts` | 191 passed, 1 skipped (192) |
  | mcp | 271/271 |

- **Docs:** `docs files --check`, `settings example --check` and `docs env --check` all exit **0**.
- **e2e**, each detached and queued:

  | Suite | Passed | Failed | Time |
  |---|---|---|---|
  | homepage, full (the new `unlisted-site.spec.ts` 3) | 97 | 0 | 3.6 min |
  | hub, full | 36 | 0 | 1.4 min (after 3.5 min in the queue) |
  | editor: `sites-crud` (the new listed round-trip 1) | 15 | 0 | 0.9 min (after 11 min in the queue) |

- **Builds:** the five above, in the proof. The editor's and umtool's `next build` were not run (no
  route or bundled path changed; tsc covers the form and the action).
- **Numbers tool:** none.
- **After the review and the merge of `main` (at `7ea35dfb`; `$T/hs-gates3.log`):** tsc clean
  (183 s, the machine under load); common 2,229/2,229 (release 15 IG's 2,220, this slice's 8 and
  the review's 1); homepage unit 12/12; homepage e2e `unlisted-site.spec.ts` 3 passed, 0 failed
  (19 s). The fixture's second list unlisting a site that carries no key, and the listed variant's
  7 sites / 39,599 transcripts, were checked by a script (`$T/hs-fixture-check2.ts`).

#### Review (verdict SHIP; `$T/hs-review.md`)

| Finding | Fix |
|---|---|
| L1: the fixture's second list relied on each site's own `listed: false` | `e3ee2eb1`: `buildFixtureInputs` writes `listed: false` on every site in it; `FixtureSite` has no `listed` |
| L2: the spec's first case would pass if the builder ignored the unlisted site | `e3ee2eb1`: it asserts the site is in the input (unlisted, its records, the shared channel) and that, listed, it adds a card and exactly its own records |
| L3: every unlisted id sorted after the listed one | `65675373`: `aaa-hidden` sorts before `beta`; the summary is still the one without it |
| L4: `unlisted` / `listed` already mean other things | `20ff9ecf`: a FACTS "Naming hazards" table of the three |
| L5: the listed site's out not searched for the own channel's name; the export site suite and `e2e:2origin` not run | left: both changes are no-ops for a site without the key, and the unit test and the real build's footer cover them |
| L6: this record's slices table, Order and Rollout did not name HS; the hub check's N | `20ff9ecf`: HS row and Order; the Rollout names HS, counts public LISTED sites and checks the new box |

#### Found and left

- **The unlisted site's own `/site.json` and `/corpus.json` still carry its `hubUrl`** (ruling 4:
  it deploys exactly as before). A visitor who adds its origin to the hub by hand gets it as any
  added origin, and the hub can read that `hubUrl` as a family member's.
- **`listed` / `unlisted` mean three things:** a video's visibility, `useHubSites`' `listed`
  flag (the hub's list has loaded), and a site the family lists. FACTS' "Naming hazards" has the
  table since the review.
- **The Rollout's hub check** ("`hub-summary.json covers N official instance(s)`") counts public
  LISTED sites; the Rollout says so since the review.
- **Unlisting takes effect at the next builds.** The homepage, the hub and every other site are
  static: until each is rebuilt and deployed, it still lists the site.
- **The editor's `/sites` list** shows no marker for an unlisted site; the form's checkbox is the
  one place (ruling 3).

#### Decisions the operator could overturn

| What I assumed | The alternative |
|---|---|
| `totals` and `availability` keep counting pool-only channels and channels of sites with no public URL, as before; only a channel exposed by unlisted sites alone leaves them | `totals` count the listed public sites only, the same scope as `official` |
| A channel an unlisted site shares with a listed site with NO public URL is not "only on unlisted sites", so `totals` count it | count a channel only when a listed PUBLIC site exposes it |
| The pooled `stats/` keep pool-only channels, as before | the pooled `stats/` hold only channels a listed public site exposes |
| `isListedSite` lives beside the key in `siteSchema.ts`, exported from `lib/site` | define it in `lib/site.ts` itself, and have the summary builder import `lib/site`'s file I/O |
| The summary's version is 6 | stay at 5 (no field changed) |
| An unlisted site's own footer still links its listed siblings | an unlisted site shows no related-sites footer |
| The checkbox sits after Public URL, with a hint naming what it removes | at the end of the form, or without a hint |

### Slice S1, as shipped — a clear screen until the first Search (2026-09-29)

Branch `r14/first-search` off `main` `99d4d76a` (`r14/two-grounds-headers` merged), worktree
`~/Projects/plans-export-header-first-search` (block #4: export e2e 3420, hub e2e 3441, editor test
3411), one Opus implementer. Scratch files `s1-*` in the job's `tmp`.

The rulings, 2026-09-29, not re-opened:
1. S1 only. The summaries download is not deferred; S2 stays a candidate.
2. A link carrying only a filter (`tg=`, `fv=`, `ch=`), like one carrying `qt=`/`q=`, runs on
   load and shows its results.

The parent's directions for the build (the plan's S1 steps):
- The gate is in `SearchResults`, which the hub mounts too, so the hub gets the same behaviour and
  `e2e:hub` joins the gate.
- The module-level `ranThisPageLife` feeds a reactive `searchedThisPageLife`; the dead
  `searchExecuted` goes. As built after the review, the flag the view reads is a second module
  flag, `askedThisPageLife`, and `ranThisPageLife` keeps `main`'s rule (review M1, below).
- A URL with `qt`, `q` or any filter key counts as asked, at hydration.
- `SearchResults` renders nothing until the flag is set; `resultGroups` is still computed. The
  page's intro (the transcript count, the Welcome card) stays.
- The bar's "Press Enter or click Search to apply" line also shows before the first Search.

My own choices are under "Decisions the operator could overturn".

| sha | what |
|---|---|
| `93d72a80` | `common:` the results area renders nothing until the visitor asks in this page life; the bar's line shows meanwhile; `searchedThisPageLife`; `searchExecuted` deleted |
| `90043b78` | `export:` `first-search.spec.ts`; one `showAll` helper, pressed where the export and hub specs read the listing at load |
| `dbdaa31b` | `plans:` this record; the plan's S1 as built; the export changelog |
| `9b01068c` | `common:` two page-life flags, one job each; a filter-only link no longer releases a stored query (review M1); the held-query comment (L1) |
| `41dc1b4b` | `export:` `restore-no-refire` covers a second mount after a filter-only link (M1) |
| _this_ | `plans:` the review's fixes recorded (L3, M1, L2 left); the S1 row of the slices table and the Rollout's checks (L4) |

- **What a plain visit shows**, on a site's `/` and on the hub's:
  - the search bar with its line, "Press Enter or click Search to apply";
  - the page's intro (on a site the transcript count and the Welcome card; on the hub the shelf,
    the figures and the archive chips);
  - the footer, whole on the first screen at 1280×800 and 390×844 (the spec asserts it).
  - No count, no Results/Chart toggle, no Copy for AI, no selection toolbar, no chart, no listing,
    no `browse-hint`, and on the hub no "N of M archives answered" line.
- **What counts as the first Search:** Enter, the Search button, a Filters Apply (all through
  `commitSearch`), a profile load or Revert (`applySnapshot`), and a URL that asks.
  - An empty Search shows "All videos (N)" and the listing, as before.
  - A filter changed before it shows nothing; the line was already there for an unapplied edit.
- **A URL that asks** (`urlAsks`, `SearchSessionContext.tsx`): any of `qt`, `q`, `tg`, the legacy
  filter keys (`ch nov nol naa nar nav nd nu m tk`) or the share-link keys
  (`fv fc ft fa fav fk fdf fdt`).
  - Presence counts, not a valid value: `?tg=Not%20An%20Id` shows every video, which is what
    `tag-chips.spec` already asserts.
  - A video (`v`, `t`), a chart's shape (`view`, `cs`), `re` and `vm` do not ask. The Share button
    always writes `fv=`, so a shared chart does run.
- **Two flags, one job each** (module state: they survive client-side navigation and reset on a
  reload):
  - `ranThisPageLife` is the hold, exactly as on `main` (release 8). It is set by a commit, a
    profile load, or an active URL query. While it is unset, a stored query is held, on this
    mount and on every later one.
  - `askedThisPageLife` is the gate: `ranThisPageLife`, or a URL that asks. `searchedThisPageLife`
    starts from it and is set wherever it is set: at hydration, in `commitSearch` and in
    `applySnapshot`.
  - So a filter-only link shows its results and runs no query. A stored query stays held on that
    load, and on a later mount in the same visit, which shows the listing under the held query.
  - On the server neither module variable is set, so the first client render matches the static
    HTML. The exported `index.html` now has the line and no results section.
- **Step 0, settled:**
  - **The hub renders `SearchResults`:** `HubHome.tsx` → `TranscriptSearch` (common) →
    `SearchResults`. So the hub has the clear screen, and `e2e:hub` ran in the gate.
  - **Back restores the listing.** Measured with a probe (not committed) on the 120-video fixture
    at 1440×1200, scrolled to 3,000 px before leaving:

    | Path | After Back | Scroll before → after |
    |---|---|---|
    | empty Search → **Use with AI** (header) → Back | the listing, "All videos (120)" | 2,838 → 2,676 px (the first card drawn: `0024` → `0025`) |
    | `qt=` link → **Use with AI** → Back | the results | 2,952 → 2,904 px |
    | empty Search or `qt=` → **Chat** (workspace nav) → Back | the listing | → 0 (the top) |
    | a card's video (the modal) → Back | leaves the page | — |

    - A route outside the workspace unmounts the session; Back remounts it with the flag already
      set, and the scroll comes back to within a card.
    - `/ask` hides the search pane in place, and the window is at the top when Back shows it again.
    - The modal opens with `replaceState`, so there is no entry of the page's own to go Back to.
    - The last two are unchanged by S1.
- **Tests:**
  - `first-search.spec.ts`, new (11):
    - on load, no results area, the line, the intro and the whole footer, at 1280×800 and 390×844;
    - Enter on an empty box shows every video, and the line goes;
    - the Search button shows every video;
    - a filter changed before the first Search shows nothing;
    - `/` → `/ask` → `/` keeps the listing;
    - Back from Use with AI keeps it;
    - a reload clears it;
    - a `qt=` link shows its results on load;
    - `tg=` and `ch=` links show theirs;
    - a restored query shows the clear screen and the filled form, and runs on Enter.
  - `showAll(page)` in `export/e2e/helpers.ts`: wait for the query builder, press Search, wait for
    `results-summary`.
  - Where specs read the listing at load:
    - `browse-all`: the first test rewritten to assert the clear screen, then the listing;
    - `charts` (7 tests), `workspace-shell` (1), `responsive` (the filters sheet and the selection
      toolbar), `tag-chips` (3; its `?tg=` tests unchanged);
    - `restore-no-refire`: the held query now shows no results area. Two new tests cover a second
      mount after a filter-only link, and each holds with no shard fetched: `?tg=` → Use with AI →
      Back, and `?ch=` → Use with AI → the header's Search link;
    - the hub's `federated-search` (8): its first test asserts the clear screen with both archives
      in, then the listing.

#### Gates (at `90043b78`; logs `$T/s1-*.log`)

- **tsc** was clean before each commit and at the tip (36–38 s). The specs commit's own run
  timed out at 100 s under the machine's load; it was run again on that commit, clean.
- **Unit:** common **2,210/2,210** (75 s). Editor unit, `test:scripts` and mcp were not run: S1
  touches no editor, mcp or scripts file, and the editor imports none of these components.
- **Builds**, each capped at 5 GB with no swap, from a clean `.next`:

  | Build | Time | Max RSS |
  |---|---|---|
  | export, site (the worktree's default) | 36 s | 957 MB |
  | export, hub | 28 s | 1,034 MB |

- **e2e**, each detached and queued:

  | Suite | Passed | Failed | Time |
  |---|---|---|---|
  | export, the specs above first | 58 | 1 | 3.4 min |
  | hub, full | 36 | 0 | 1.7 min |
  | export, full | 254 | 0 | 11.4 min |
  | `e2e:2origin` (`E2E_TWO_ORIGIN_REBUILD=1`, the `export/public` links in place) | 3 | 0 | 49 s |

  The one failure in the first run was the new `/` → `/ask` test: that run's first visit to `/ask`,
  compiled by the dev server, took longer than 5 s. The test now waits 20 s for the URL. After
  `e2e:2origin`, the primary's `export/public` files kept their 2026-09-28 mtimes.
- **Numbers tool:** none.

#### Found and left

- **Back from `/ask`** shows the listing at the top, not where the reader left it (the table
  above). This is as before S1.
- **A card's video opens without a history entry**, so Back from the modal leaves the site. This is
  as before S1.
- **Enter before the session hydrates (review L2).** Until the manifest is in, a `qt=` link shows the
  line and not yet its results. Enter in that window commits the empty draft, and `commitSearch`
  strips `qt` before hydration reads it. The race is on `main` too; the line now shows during it.
  Gating the line's first-Search half on hydration would close it, at the cost of the line in the
  static HTML. Left for the operator.

#### Decisions the operator could overturn

| What I assumed | The alternative |
|---|---|
| The line also shows under the bar on a fresh `/ask` (the bar is shared, and Search there applies the grounding) | show it only on the search view |
| A link with only a chart's shape (`view=chart&cs=…`, copied from the address bar after an empty Search) opens on the clear screen | count `view`/`cs` as asking |
| A video link (`?v=`) opens the video over the clear screen | count `v` as asking |
| On the hub, the "N of M archives answered" line waits for the first Search with the rest of the results; each failed archive's chip says so, with its Retry, before that | show the line above the gate |
| After a filter-only link, a later mount in the same visit (a page outside the workspace and back, the header's Search link) shows the listing, with the stored query held in the box: the gate is once per page life | the clear screen again on that mount |
| A key's presence asks, not a valid value (`?tg=Not%20An%20Id`) | count only the keys the page applied |
| The Search button keeps its outline look before the first Search; only the line says what to do | fill it as for an unapplied edit |

#### Review (verdict SHIP AFTER FIXES; `$T/s1-review.md`)

The reviewer also ran the editor suite's specs that drive the export's bar, results and `?v=`
modal (`export-search`, `export-player-platform-cache`): 21 passed, 0 failed.

| Finding | Fix |
|---|---|
| M1: a filter-only link set `ranThisPageLife`, so on a later mount in the same visit (a page outside the workspace and Back; the hub's `/` → `/ask`) a query stored by an earlier visit ran and fetched shards, where `main` held it. The reviewer's probe: `?tg=` → Use with AI → Back fetched 1 shard and showed "Matching videos"; `?ch=` → Use with AI → home did the same | `9b01068c`: two module flags. `ranThisPageLife` is set at hydration only by an active URL query, as on `main`, and `askedThisPageLife` (a run, or `urlAsks`) drives the gate. `41dc1b4b`: `restore-no-refire` seeds a stored query and checks both paths: 0 shards, the box filled, Search marked, and "All videos" |
| L1: the held-query comment still said "the browse listing" | `9b01068c` |
| L2: Enter before hydration on a `qt=` link commits the empty draft (a race on `main`; the line now shows during it) | left: "Found and left" |
| L3: the rulings list mixed the rulings with the plan's steps and my choices | this commit: the rulings, the parent's directions, and my decisions, apart |
| L4: the slices table's S1 row, and the Rollout's live checks | this commit: the row; three S1 checks under "Live checks" |

#### Gates after the fixes (at `41dc1b4b`; logs `$T/s1-*2.log`, `$T/s1-fix-e2e.log`)

The fix changes only which flag each place sets, so the full suites were not re-run.
- **tsc** clean before each fix commit (182 s under the machine's load).
- **common** 2,210/2,210.
- **e2e**, detached and queued (spec lists in `$T/s1-fix-specs.txt`, `$T/s1-fix-hub-specs.txt`):

  | Suite | Specs | Passed | Failed | Time |
  |---|---|---|---|---|
  | export | `first-search`, `restore-no-refire`, `browse-all`, `workspace-shell` | 20 | 0 | 5.1 min |
  | hub | `federated-search` | 12 | 1 | 2.5 min |
  | hub | `federated-search`, again | — | — | the dev server did not answer in 120 s |
  | hub | `federated-search`, again | 13 | 0 | 1.7 min |

  The hub's first run failed on its first test's first line, the archive chip still "loading"
  after 5 s. That line is unchanged from `main` and comes before anything S1 changed; the machine's
  load average was 33. On the second attempt the dev server did not start within 120 s. The third
  run passed 13/13.

### Slice CF, as shipped — the growth chart folds the small sites into Other (2026-09-29)

Branch `r14/chart-fold` off `main` `69e058d6` (slice HS merged), worktree
`~/Projects/homepage-social-visible` (block #3: homepage e2e 3340, homepage static 3331), one Opus
implementer. Scratch files `cf-*` in the job's `tmp`. The rulings (2026-09-29, not re-opened):
1. A site under **5 %** of the placed total — the chart's total over the whole plotted range, the
   sum the bands are placed from — folds into ONE "Other" band, drawn on top of the stack, in a
   neutral grey from the chart tokens that keeps ≥ 3:1 on the Light and the Dark ground.
2. A fold of one is no fold: with a single site under the threshold, nothing folds.
3. The legend shows the kept sites plus "Other"; the hover title and the "Numbers by year" table
   still name every site.
4. Colour follows the site, never its rank: a kept site's colour does not change because another
   folded.
5. The instance cards and `/stats` are unchanged; the surface gap draws correctly with Other on
   top.
6. (The review, ruled by the parent.) Other wears a dedicated `--chart-other` token in both chart
   blocks of `tokens.css`: Light `#3e545c`, Dark `#62625c` (the near-neutral).

| sha | what |
|---|---|
| `8351d24f` | `homepage:` the fold in `lib/growthGaps.ts` (`foldedSites`, `growthLayers`, `ownLayers`, `layerColors`; `growthStack` stacks layers); the chart draws the layers, its legend, label and caption; the unit tests; the e2e fixture's fifth site at 4 a day |
| `6e0b652d` | `homepage:` e2e: `growth-chart.spec.ts` (the fold, the grey), `instance-colours.spec.ts`, `marketing.spec.ts` adjusted |
| `e95a8058` | `plans:` this record, the slices table, Order and Rollout; FACTS; the homepage changelog |
| `17832633` | `common, homepage:` `--chart-other` in both chart blocks with its numbers; `OTHER_COLOR` points at it; the chart's header comment; `growth-chart.spec.ts` resolves `OTHER_COLOR` and pins each base's value (review M1) |
| `c42d2bbc` | `homepage:` `instance-colours.spec.ts` reads each site line's stroke on `/stats/`, the Vermilion site's the rust (review L2) |
| `afddd66e` | `plans:` the review's fixes in this record (the grey, the Review table, the screenshots, Found and left, Decisions); the CF row's files; FACTS; the homepage changelog |
| `9267ec08` | merge of `main` `721ed0eb` (slice S1); `plans/release-14.md` resolved by hand (below), the changelogs clean |
| _this_ | `plans:` the post-merge gates; the commit table |

**What shipped.**
- **The rule** (`homepage/app/lib/growthGaps.ts`, pure, beside `growthStack`):
  - `foldedSites(months, sites)`: each site's sum over the plotted months against the sum of all of
    them. A site is under when `100 × its sum < FOLD_PERCENT × the total` (`FOLD_PERCENT = 5`), on
    integers, so exactly 5 % keeps its band. The folded sites are returned only when two or more
    are under; none when the total is 0.
  - `growthLayers(months, sites)`: the kept sites in stack order (the summary's, `STACK_ORDER`
    unchanged), then `{ key: "(other)", sites: [the folded indexes], other: true }` — a key no site
    id can be (`[a-z0-9][a-z0-9-]*`). `ownLayers(sites)` is every site its own band.
  - `growthStack(months, sites, layers = ownLayers(sites))`: one band per layer, a layer's value
    in a month the sum of its sites'. It returns `layers` in place of `order` (the chart was
    `order`'s only reader). Totals, peak, value scale and gridlines are over every site, so the
    stack's top and the scale are the same folded or not.
  - `layerColors(layers, sites)`: `siteChartColors` over EVERY site, and a kept site takes its own
    entry, so a fold never repaints one (with no accents, the kept list alone would move a site
    after a folded one to a lower slot; the unit test proves it); Other is `OTHER_COLOR`.
- **The grey is its own token, `--chart-other`** (`common/styles/tokens.css`, beside `--chart-axis`
  in both chart blocks, each with its numbers in a comment as `--chart-6` has; review M1): Light
  `#3e545c` (OKLCH L 0.43, C 0.03), Dark `#62625c` (L 0.49, C 0.01). Contrast: 7.36:1 on the Light
  ground and 7.99:1 on its chart surface; 3.22:1 on the Dark ground and 3.06:1 on its chart surface
  (the Dark slots are 3.37–6.46:1 on the ground: Other is the dimmest mark there). Against each
  chart slot it can sit on (the dataviz skill's validator's measures, OKLab ΔE × 100, normal / the
  worse of protan and deutan):

  | Slot | Light | Dark |
  |---|---|---|
  | blue (`--chart-1`) | 21.9 / 21.4 | 17.8 / 17.9 |
  | green (`--chart-2`) | 17.4 / 15.2 | 19.2 / 15.7 |
  | violet (`--chart-3`) | 16.2 / 13.4 | 23.4 / 21.5 |
  | amber (`--chart-4`) | 22.1 / 18.2 | 20.3 / 18.1 |
  | magenta (`--chart-5`) | 24.5 / 9.1 | 19.5 / 7.3 |
  | rust (`--chart-6`) | 14.1 / 10.5 | 13.5 / 9.3 |

  - Every CVD pair clears the floor (6) on both bases, and the target (8) on Light. Dark's magenta
    (7.3) is in the 6–8 floor band, legal with the legend, the table and Other's place on top.
    Rust is the one pair under the normal-vision 15, on both bases.
  - A grey is under the validator's chroma floor by definition (it is the de-emphasis role, not a
    categorical slot).
  - Other touches only the kept site beneath it (the highest with a height that month). In
    today's data that is Bonnellyzer, blue, in every month Other has data.
  - The first build used `--chart-axis` (Light `#55646e`, Dark `#8a8170`: magenta at CVD 4.5 /
    4.2, and on Dark green at 4.1, under the floor). The review's sweep found in-band greys that
    clear CVD 8 against every slot; the parent ruled the token above.
- **The chart** (`ArchiveGrowthChart.tsx`):
  - the legend is the layers: the kept sites' titles, then "Other";
  - the areas and the gaps are keyed by layer;
  - the hover title is unchanged: every site with data that month, by name, folded or not;
  - the table is unchanged: a column per site;
  - the image's `aria-label` gains, when something folds, "Hasanalyzer, Rekietalyzer and
    Jasolyzer, each under 5% of the total, are drawn together as Other." (the legend is hidden
    from assistive tech);
  - the caption gains, when something folds, "Instances under 5% of the total are drawn together as
    Other.";
  - the header comment carries the grey's numbers.
- **The gaps are unchanged code.** They work on bands, so Other is one band, the top one:
  `bandAbove` of the top kept site is Other wherever Other has a height, and Other's own upper edge
  has no gap (nothing sits on it). On today's summary the gap segments are the same folded and
  unfolded, and none is drawn under Other: Bonnellyzer is under 3 px wherever Other sits on it, so
  they touch, as the three small bands did before.

  | Plot height | Segments | Runs | Under Other |
  |---|---|---|---|
  | 200 px | 73 | 7 | 0 |
  | 260 px | 112 | 6 | 0 |
  | 300 px | 119 | 7 | 0 |

  On the fixture, gaps are drawn under Other, and no band is covered at any height (unit test).
- **Unchanged:** the instance cards, `/stats`, `siteChartColors`, the hub; in `tokens.css` only
  `--chart-other` is added.
- **With today's data** (the primary's `homepage-summary.json` of 2026-09-28, 75,821 transcripts
  on the chart): Jeralyzer 42.03 %, Anilyzer 38.40 %, Bonnellyzer 9.18 % keep their bands;
  Hasanalyzer 4.33 %, Rekietalyzer 3.79 % and Jasolyzer 2.26 % are Other.

**The e2e fixture** (`homepage/e2e/fixture-summary.ts`). Its six sites were 60.21, 12.90, 9.68,
8.60, 5.38 and 3.23 % of the chart: one under 5 %, no fold. The smallest change that gives two:
`fixture-five` transcribes 4 a day, not 5 (1,600 recordings, not 2,000). The shares are now 60.87,
13.04, 9.78, 8.70, **4.35** and **3.26 %**, and `fixture-five` and `fixture-six` fold. The
unlisted site stays out of the chart (it is out of the summary). Every changed expectation:
- The fixture's listed totals: 36,799 transcripts, not 37,199; with the unlisted site listed,
  39,199, not 39,599 (HS's record has the old pair). No spec reads either: `unlisted-site.spec.ts`
  computes both sides.
- `growth-chart.spec.ts`, the pixel test: one colour per band (four kept sites and Other), not per
  site.
- `instance-colours.spec.ts`: the legend has five swatches, not six (the four kept sites in their
  slots among all six, then Other's grey); `fixture-five`'s card is no longer compared with a
  legend swatch (it has none); `fixture-six`'s hue check reads its chart colour, the rust, not a
  legend swatch.
- `marketing.spec.ts`: the caption's check is `/all official instances\.( |$)/i`, not `/…\.$/i`,
  since the Other sentence can follow.

**Tests.**
- `growthGaps.test.ts`: the fixture test now runs the chart's layers (the fold is `[4, 5]`; a gap is
  drawn between the top kept site and Other; none along Other's top; no band covered at any
  height, folded or not). New:
  - today's proportions (42.03 / 38.40 / 9.18 / 4.33 / 3.79 / 2.26 %, the family's accents): the
    three largest keep their bands, Other holds the other three, its top is every month's total,
    the colours are amber, magenta, blue and the grey;
  - the edge: 999 of 20,000 (4.995 %) folds, 1,000 (exactly 5 %) does not; two sites at exactly
    5 % fold nothing;
  - a single site under 5 %: no fold;
  - no site under: `growthLayers` is `ownLayers`, and the stack is `growthStack`'s default;
  - all but one under: one kept band and Other;
  - a site with nothing in the range folds with another; nothing plotted folds nothing;
  - 25 equal sites: every site folds (see "Decisions");
  - colour stability: the kept sites' colours equal the unfolded run's, and differ from what the
    kept list alone would give.
- `growth-chart.spec.ts`, new:
  - two sites under 5 % are one Other band on top: the layers are the four kept sites and
    `(other)`; the legend reads the four titles and "Other"; the areas' fills, in paint order, are
    the four slots and `OTHER_COLOR` (`var(--chart-other)`) last; the label names "Fixture Five and Fixture Six";
    the caption says what Other is; every month's title names every site with data that month and
    no title says "Other"; the table's header is Year, the six titles, Total;
  - Other is `OTHER_COLOR` on both grounds, rendered `rgb(62, 84, 92)` on Light and `rgb(98, 98, 92)`
    on Dark (so each base declares its own), not the axis colour, at least 3:1 on the ground, and no
    kept site's colour (review M1).
- `instance-colours.spec.ts`, new (review L2): on `/stats/`, on both bases, the six site lines'
  rendered strokes are the six sites' chart colours in the summary's order; the Vermilion site's is
  the rust (`--chart-6`) and within 25° of Vermilion's hue. The card test's own site-5 check is the
  card's hue against its chart colour; the tautology it replaced (`chart[5]` against
  `var(--chart-6)` twice) is gone.

**They bite** (each change made by hand, the unit tests run, the change reverted):
- `<=` for `<` in the rule: the edge test fails.
- A fold of one allowed: the single-site test fails.
- Colours from `siteChartColors` over the kept sites alone: the colour-stability test fails (the
  today's-proportions test does not: its sites' slots come from their accents).
- Other at the bottom of the stack: four tests fail (the fixture's, today's, all-but-one, colour).

#### Gates (at `6e0b652d`; logs `$T/cf-*.log`)

- **tsc** clean in every package on the tree of `6e0b652d`, run before the first commit (79 s);
  `8351d24f`'s tree has `main`'s versions of the three spec files, which import nothing it changed.
- **Unit:** common **2,229/2,229**; homepage **20/20** (12 before, 8 new).
- **Build**, capped at 5 GB with no swap, from a clean `.next`, with the primary's summary copied
  into the worktree's `homepage/public` for the screenshots (the worktree's own put back after):
  `pnpm --filter homepage exec next build` **ok**, 39 s, max RSS 779,504 KB. `main`'s chart,
  built the same way for the "before" shots: 80 s, 754,160 KB. After the review (`17832633`, for
  the retaken shots): 20 s, 809,752 KB.
- **Homepage e2e, full** (97 at `main` after HS; 2 new):

  | Run | Passed | Failed | Time |
  |---|---|---|---|
  | first (`$T/cf-e2e-homepage.log`, after 9.5 min in the queue) | 98 | 1 | 4.2 min |
  | again (`$T/cf-e2e-homepage2.log`, after 5.5 min in the queue) | **99** | **0** | 3.9 min |

  The first run's failure was `marketing.spec.ts`' "Changelog is reachable from the footer on
  every page": the 30 s test timeout, reached on the seventh of its seven pages (the dev server
  compiling each on first visit, the machine's load average 11–17 with other suites running). It
  passed in 5.3 s in the second run and in the three-spec run below; nothing in it reads the
  chart.
- **Along the way:** `growth-chart`, `instance-colours` and `marketing` specs, 21 passed, 0 failed
  (1.8 min).
- **Numbers tool:** none.
- Not run: the export, hub and editor suites and their builds (no file of theirs changed).

**Screenshots** (`~/reports/release-14/shots/chart3/`, 2×, the static server on 3331, today's
summary):
- `after-{390,1280}-{light,dark}.png`: the chart with `--chart-other` (retaken after the review;
  the operator judges the Dark one); `…-2019-2026.png`: the plot from 2019 on, where Other lies;
- `after-axis-…`: the same from the first build, Other in `--chart-axis`, for comparison;
- `before-…`: the same from `main`'s chart and the same summary (six bands);
- `after-1280-{light,dark}-table.png`: "Numbers by year" open, a column per site.

**Found and left:**
- **The pairs short of the validator's targets** are Dark's magenta (CVD 7.3, the 6–8 floor band)
  and rust on both bases (normal 14.1 / 13.5). They matter only when that site is the top kept one under Other; today it
  is blue.
- **`/stats`' channel breakdown has an "Other (N)" of its own**, in `--muted-foreground`
  (`common/lib/homepageChart.ts` `OTHER_COLOR`), not `--chart-other`. `/stats` is unchanged, as
  ruled.
- **`--chart-other` is not in `REQUIRED_TOKENS`** (`common/components/themeConfig.ts`, the list
  `themeTokens.test.ts` checks every base declares): that file is outside this slice.
  `growth-chart.spec.ts` pins each base's rendered value instead (a Dark block without it would
  paint the Light value from `:root`).
- **The fold is the homepage chart's only.** `/stats`, the hub and the sites' charts draw every
  site, as ruled.

#### Decisions the operator could overturn

| What I assumed | The alternative |
|---|---|
| The caption gains one sentence saying what Other is, only when something folds | the caption unchanged |
| The image's label names the folded sites | name only how many |
| The legend reads "Other", with no count or names | "Other (3)" |
| The hover title lists every site with data in the summary's order, with no Other subtotal | an "Other N" line, its sites under it |
| When every site is under 5 % (21 or more sites), every site folds: one Other band | keep the largest, or fold nothing |
| A site with nothing in the plotted range is under 5 % and folds with another | leave it out of the chart |

#### Review (verdict SHIP AFTER FIXES; `$T/cf-review.md`)

| Finding | Fix |
|---|---|
| M1: the axis grey was under the CVD floor against magenta on both bases (4.5 / 4.2) and against green on Dark (4.1); the record's sweep sentence overstated the case against a better grey | `17832633`: `--chart-other` in both chart blocks (Light `#3e545c`, Dark `#62625c`, as ruled), `OTHER_COLOR` points at it, the chart's comment and the spec follow; the sweep sentence is withdrawn and the token's measured numbers stated ("What shipped"); the labels' shared colour is gone from Found and left; the four `after-*` shots retaken, the first build's kept as `after-axis-*` |
| L1: the CF row omitted `plans/FACTS.md` | `afddd66e`: the row lists it, and `common/styles/tokens.css` |
| L2: `instance-colours.spec.ts` checked the sixth site's colour against itself | `c42d2bbc`: a rendered check of every site line's stroke on `/stats/`, the sixth the rust |
| L3: "Found and left" left rust out of the weak pairs | moot with M1; the line names Dark's magenta and rust |
| L4: the merge of `main` (`721ed0eb`, S1 merged) conflicts in this record only | `9267ec08`: S1's row, then CF's; S1's section, then CF's, before "## Rollout"; the Order line and the Rollout's intro name S1 and CF; both sets of live checks. The changelogs merged clean, every bullet under `[Unreleased]` (checked by eye) |
| L5: the changelog's "at this release's numbers" will drift | left, as a release note |

The caption and the image-label sentences are kept (the review: acceptable additions).

**Gates after the review and the merge of `main`** (at `9267ec08`; logs `$T/cf-tsc2.log`,
`$T/cf-tsc3.log`, `$T/cf-gates2.log`, `$T/cf-e2e3.log`):
- **tsc** clean in every package before the fix commits (44 s) and on the merged tree before the
  merge commit (164 s).
- **Unit:** common **2,229/2,229**; homepage **20/20**.
- **e2e**, detached and queued: `growth-chart`, `instance-colours` and `marketing`, **22 passed, 0
  failed** (1.1 min; the new `/stats/` line test among them). The full suite was not re-run, as
  directed.
- **Build** (for the retaken shots, at `17832633`): 20 s, 809,752 KB, capped.

## Rollout

Release 14 is slice HP (merged, `bfa1ff3c`), `r14/two-grounds-headers`, slice HS
(`r14/hidden-sites`), slice S1 (`r14/first-search`) and slice CF (`r14/chart-fold`), each after the
parent's merge. Every command below is typed **from the primary checkout's root**. There is no `archilyzer` on PATH, so it is `pnpm archilyzer …`. The
command forms are the ones verified in `plans/stats-cache-key.md`'s rollout.

**Preconditions.**
1. `main` carries `r14/two-grounds-headers`.
2. **The homepage build runs release 12's source publish.** So the homepage waits on release 12's
   rollout step 0: the denylist is complete and `pnpm archilyzer source publish --check` is clean.
3. **No other index, stats or site build is running.** `/jobs` shows none running or queued.
4. **If the stats-cache-key rollout has not run yet, run it first.** Its steps 5–7 build the
   homepage, the hub and the sites in the same order as below, and then cover this release's too.
5. **The link to keep in the header on small screens is marked, before any build.** The parent
   marks it in the editor's Settings (**Keep in header on small screens**) and saves. The builds
   are static: a link marked after them shows only after another build. With none marked, no
   header shows a social link below 520 px.

**The order, and why.** The homepage goes first, because every site's header links to
`https://archilyzer.pages.dev/#instances` and that anchor exists only in the new homepage. The
hub goes before the sites, because in basic mode they share `export/out`. The six sites go last.

1. **The editor**, for its own header (no theme menu; the toggle cycles System, Light and Dark):
   `pnpm --filter editor build`, then restart :3001 the way it is normally run.
2. **The homepage.** Use exactly ONE of:
   - `pnpm archilyzer build homepage && pnpm archilyzer deploy homepage`;
   - `pnpm ops build-homepage --json '{"deploy":true}' --wait`.
3. **The hub, before the sites.** Build it, check its log, and only then deploy it. Use exactly
   ONE pair:
   - `pnpm archilyzer build hub`, then `pnpm archilyzer deploy hub`;
   - `pnpm ops build-hub --wait`, then `pnpm ops deploy-hub --wait`.

   **Between the two,** the build's `compose-hub: …` line must end with `hub-summary.json covers
   N official instance(s)`, where N is the number of public LISTED sites (a site whose
   **List on the Archilyzer homepage and hub** is unticked is not counted). If it says
   `hub-summary.json skipped: …`, stop and fix what it names.
4. **The six sites:** `pnpm ops build-deploy --json '{"all":true}' --wait`.

**Live checks.**
- `https://archilyzer.pages.dev/#instances` lands on Official Instances, below the header.
- One site at 390 px wide:
  - the header shows the mark and the site's name, then the marked link and the theme toggle as
    one group, then the menu button;
  - the menu holds the nav and "Archilyzer".
- The same site at 520 px wide and more: every social link, up to four.
- At 1280 px wide, the header's **Archilyzer** link goes to the homepage's Official Instances, in
  the same tab. There is no Sites menu and no Hub link.
- **Changelog** is in the footer, after Use with AI, and not in the header.
- **A stored Sepia renders Light.** In a site's console, run
  `localStorage.setItem("ytdlp-tb:base","sepia")` and reload. `<html data-base>` reads `light`,
  and `localStorage.getItem("ytdlp-tb:base")` now reads `"light"`.
- The homepage's growth chart has no slash in the page colour through any band. `/stats` in
  Area mode has coloured top lines.
- The growth chart's legend lists the sites with 5 % or more of the chart's total, then
  **Other** (a grey band on top); the caption ends "Instances under 5% of the total are drawn
  together as Other."; **Numbers by year** has a column for every site. `/stats` still has a line
  per site.
- Every site's settings in the editor show **List on the Archilyzer homepage and hub**, ticked. With
  none unticked, the homepage, the hub and every footer list the same sites as before, and
  `https://archilyzer.pages.dev/homepage-summary.json` reads `"version":6`.
- **S1, a clear screen until the first Search**, is in every site's build and the hub's, so steps 3
  and 4 above (every site and the hub rebuilt and deployed) carry it. On one site:
  - a plain visit to `/` shows the search bar, the transcript count and the whole footer, with no
    scrolling and no listing;
  - a `qt=` link shows its results on load;
  - on a site that publishes tags (Anilyzer): Search for a word, then open a `tg=` link in the
    same tab. Its listing shows with the word held in the box. After another page and Back the
    word is still held: the box is filled, the Search button is marked, and the results say "All
    videos", not "Matching videos".