@nanisoft/prism-llms 0.4.0 → 0.5.0

package/dist/md/changelogs/prism-tokens.md

@@ -0,0 +1,126 @@
+# @nanisoft/prism-tokens
+
+## 0.6.0
+
+### Minor Changes
+
+- 4f9ce77: A server-rendered pack boundary can now be mode-correct
+  
+  The only dark selector for a pack was `[data-pack="<id>"].dark`, which requires
+  the mode class on the same element. A server cannot know the reader's mode, so a
+  subtree carrying a second pack rendered in that pack's light values on a dark
+  page. The token build now emits the two-member list
+  
+  ```
+  [data-pack="<id>"].dark, .dark [data-pack="<id>"]
+  ```
+  
+  as one rule with one declaration block, for all five packs. The compound half is
+  retained, because it is what a dark surface on a light page is and it is a
+  capability the package already published.
+  
+  Consumers get a second change alongside it. `themeAttributes` now takes `mode` as
+  optional, because a required one is that same impossibility written into the type:
+  the only way to ask for a boundary with no mode class of its own was to pass
+  `'light'` as a guess, and the guess is wrong for half of readers. Omitting `mode`
+  is the honest spelling of "wear the mode of the element carrying `.dark`". The
+  change is a widening, so every existing caller still compiles.
+  
+  The theme-boundary law in `DESIGN.md` is rewritten to match. It previously
+  instructed an implementer to put `data-pack` and `class="dark"` on a boundary
+  together, which works for a client-applied boundary and cannot work for a
+  server-rendered one.
+- 5dcb347: Every colour role now has a contrast row or a stated reason, and there is a brand ink
+  
+  The contrast gate decided what to measure by a naming convention: any root token ending in
+  `-foreground` had to be the first element of a pair. Six of the seven roles that were in no row at
+  all could not be reported by that rule, because `chart-1` through `chart-5` and `sidebar-border`
+  do not end in `-foreground`, and `sidebar-ring` does not either. A new role therefore defaulted to
+  silence. The gate now walks the token source instead of the table, so a role nobody has written
+  down is a failure rather than an absence, and the role set is read from `src/semantic/` rather than
+  from the pair table, or the walk would be checking the table against itself.
+  
+  `sidebar-ring` is what the walk immediately found. It had no row at all, and its shipped values
+  measured 2.42:1 in the base pack's light mode, 1.73:1 in its dark mode, 2.86:1 in Mint and 2.97:1
+  in Sky against the sidebar surface: a focus indicator four of twelve pack and mode combinations
+  could not be seen against, reported as compliant because it was not measured. The values moved, not
+  the row. The base pack goes from neutral 400 and neutral 700 to neutral 500 in both modes, because
+  the base pack's ring is one ramp rather than two; the pastels go from brand 500 to brand 600 in
+  light, because the sidebar is neutral 50 rather than neutral 0 and brand 500 is one step short of
+  3:1 in two packs. The dark pastels keep brand 300 at 8.06:1 or better. `sidebar-border` gains an
+  advisory row on the sidebar surface with the same judgement `border` and `input` already carry,
+  which is how its 1.18:1 to 1.37:1 became a number on the record rather than an unmeasured value.
+  
+  `chart-1` through `chart-5` are exempted with a reason rather than measured against a ground. A
+  series is a graphic object: shadcn consumes the five as the stroke and fill of a series and reads
+  its tooltip and legend text from `foreground` and `muted-foreground`, which are each gated against
+  their own ground. What binds a five-way set is that no two series resolve to the same value, so
+  that is what the gate asserts instead, in every pack and both modes. The measured tightest pair is
+  37/255 in a channel, Peach in dark mode between `chart-1` and `chart-5`. No minimum separation is
+  asserted, because a threshold with no standard behind it is a number that gets bent. `radius` is
+  exempted as a length rather than a colour, and the exclusion is a declared entry with a reason
+  instead of a name filter buried in a loop. Every exemption is printed on every run with the roles
+  that share it.
+  
+  **`brand-ink` is a new semantic role: a brand hue read as text.** Three sites want a brand colour
+  they can read on a surface and `primary` cannot serve it, because `primary` is a fill: a pastel
+  brand 400 measures 2.20:1 to 2.44:1 on its own pack's page and 1.90:1 to 2.05:1 on the accent
+  surface. It is brand 700 in light and brand 200 in dark, in all five packs and the base, and it is
+  gated at 4.5:1 against three grounds rather than one: the page, a card, and the accent surface.
+  The accent is the ground that decides the step in both directions, because a light ink has to
+  survive the pack's own brand 100 and a dark ink has to survive brand 800, and brand 600 and brand
+  400 respectively are the steps that fail each. The measured worst case across all six sources, both
+  modes and all three grounds is 5.57:1 in Mint's light mode. The dark step is brand 200 rather than
+  the brand 300 the accent ground would have permitted, because brand 300 is `primary` in every
+  pack's dark mode and a brand ink that is the brand fill is the confusion the role exists to end. A
+  test asserts that `brand-ink` never resolves to the same value as `primary`, `primary-foreground`
+  or `foreground` in any source or mode.
+  
+  **The two values that were already in the contract are the two answers this gate rules out, and
+  both are written into `DESIGN.md` as prohibitions with a number attached.** `primary-foreground` is
+  the pack's brand 950 step, in both modes and all five packs, published under three further names in
+  light mode, and it is the value a reader of the contract reaches for when they want the brand colour
+  as text. It clears its own row by 7.39:1 to 11.06:1 and it measures 1.00:1 to 1.82:1 on the dark
+  page ground, the dark card and the dark accent surface. In light mode it reads 15.05:1 to 18.05:1
+  on the page, so half the modes it looks right and half it disappears, and no row in the contract
+  could see the difference. `ring` is the other: the pack's brand hue as a non-text boundary, gated
+  at 3:1 against the page ground alone, and it measures 2.66:1 to 4.35:1 on the accent surface in all
+  six light-mode sources. Neither was wrong arithmetically. Both are invisible on the ground a brand
+  ink lands on, and both now have a role that is measured there.
+  
+  **A role is measured in both modes or the run fails.** Every colour role, exempt ones included, has
+  to resolve to an sRGB value in light and in dark in every pack, so an exemption buys a role freedom
+  from a ground and never from a mode. This is what a role present in one mode and absent from the
+  other looks like, and the build's own key-set guard does not reach it because it compares light
+  against light.
+  
+  The old `-foreground` name test is gone rather than kept beside the walk. It was never a second
+  opinion on coverage, which the walk settles, and its only additional content is that a role named
+  `*-foreground` must be the foreground of a row rather than the background of one, which is a claim
+  about the shadcn naming convention and belongs to the Contract Rule.
+  
+  No consumer action is required. `brand-ink` is additive: it is a new custom property, the stylesheet
+  gains a `text-brand-ink` utility, and no existing token changed except `sidebar-ring`, whose value
+  moved in the base pack and in the light mode of every pastel. A consumer that styled its own sidebar
+  focus ring from `--sidebar-ring` will see a more visible ring, which is the point.
+
+## 0.5.1
+
+### Patch Changes
+
+- 2fc6a0a: Retune `muted-foreground` so muted text clears 4.5:1 on the muted surface
+  
+  The base pack's `muted-foreground` moves from neutral 500 to neutral 600. It
+  cleared 4.5:1 on the page background but fell to 4.34:1 on `muted`, where the
+  pill, avatar fallback, kbd and tab-list pattern place it. The contrast gate now
+  also checks `muted-foreground` on `muted`. The pastel packs were already pinned
+  to their neutral 700 and are unchanged.
+
+## 0.5.0
+
+Prism was rebuilt from a fresh repository on a clean break. The token pipeline is
+DTCG 2025.10 compiled with Style Dictionary, the semantic names keep the shadcn
+contract verbatim, and motion, typography, spacing, shadows, breakpoints and
+containers all reach CSS through a single `@theme static` block.
+
+Changesets are appended above this entry for every published change.

package/dist/md/changelogs/prism-ui.md

@@ -0,0 +1,429 @@
+# @nanisoft/prism-ui
+
+## 0.6.0
+
+### Minor Changes
+
+- 956fc8f: A call to action with a destination renders a link, and an icon is required only where a tile is
+  
+  Every primary call to action in the Block family declared an `href` its type then
+  never rendered. The control was a `<button>` with no handler: the type said the
+  component could navigate, the documentation implied it, and the markup could not
+  do either. Sixteen of them across the family, five of them to another origin.
+  
+  **`CtaLink` is a new Component.** It renders a native `<a>` and carries the
+  visual weight of a call to action, so assistive technology announces a link and
+  the browser's own affordances all work: a status bar showing the destination, a
+  context menu to copy it, middle-click to open a new tab. It is a Component in its
+  own right rather than an `as` or `href` prop on `Button`, which is the answer
+  `BreadcrumbLink` and `PaginationLink` already give twice in this package. It is a
+  server component, so the swap costs no client JavaScript.
+  
+  `href` is required on `CtaLink`: an anchor without one is not a link. `newTab` is
+  a declared prop and the only way to open a new browsing context, because `target`
+  is not accepted, so the relationship cannot be opted out of by accident. When it
+  is set, `rel` defaults to `noopener noreferrer`; a consumer that needs a
+  different relationship passes `rel` and it wins. Prism does not parse the
+  destination and does not decide what counts as external.
+  
+  **`Cta01` no longer offers an action that navigates nothing.** `Cta01Action`'s
+  `href` is required, there is no optional arm and no dead-button arm, and a new
+  `actionSlot` takes a control the Block does not own, such as a client router's
+  link. The Block makes no guess about cross-origin. The one rendered change in
+  this release is a button becoming the link its type always said it was: the link
+  is styled from the same recipe as the button it replaces, at every shared variant
+  and size, and a test asserts that class-for-class.
+  
+  **`FeatureGrid01`'s icon is required only where a tile is rendered.** The
+  requirement is a union discriminated on `variant` rather than an optional prop,
+  because an optional prop turns a safe-at-render condition into an unsafe one and
+  permits an empty accent tile per feature. The `icon` variant is the one an
+  omitted `variant` selects, so the default and the required icon live on the same
+  arm; the `bare` variant renders no tile and requires no icon, and the type says
+  so rather than the renderer finding out. `IconFeature` and `BareFeature` are
+  exported, and the union is the type the site's API table is extracted from.
+  
+  **The version the four downstream site plans pin is `0.6.0`.** It is
+  `@nanisoft/prism-ui@0.6.0`, and `@nanisoft/prism-tokens@0.6.0` with it, because the
+  two are linked in `.changeset/config.json` and the pending
+  `a-server-rendered-pack-boundary-can-be-mode-correct` changeset already declares a
+  minor for both from `0.5.1`. This one declares a minor as well and so does not
+  raise it.
+  
+  Migration: a `Cta01` action needs a destination. Where the action was a label
+  alone, give it the `href` it was always declared with, or move the control into
+  the new `actionSlot`. A `FeatureGrid01` in the `bare` variant can now drop the
+  `icon` its type used to demand; the `icon` variant still demands one per feature.
+- d001983: A diagram is a Component, and a product's mark is one item rather than two copies
+  
+  Three of the four product sites draw something in the position where an empty box
+  is most visible, and all of it was a canvas. A canvas reads computed style from
+  one element while a pack is an attribute on an ancestor, so a scoped `data-pack`
+  boundary hands it the pack's light values on a dark page, and no token gate can
+  see it: the value the drawing holds is a legal token value, applied to the wrong
+  pack. Cutting the canvases without a replacement is the alternative, and it
+  leaves three heroes with an empty slot.
+  
+  `Diagram` draws named things and the labelled relations between them as inline
+  vector markup. Every stroke and every fill is a semantic utility, never a
+  resolved value, so a boundary restyles the drawing through the cascade exactly as
+  it restyles a heading. It is a server Component: no hook, no mode, no context, no
+  prop spread, and no client code, so a consumer renders it from a server file
+  with no provider mounted. Its node mark is a circle and a relation is a path
+  because both are shapes with no radius concept, which is what keeps a boundary
+  from moving anything but colour; a `<rect>` was rejected because its corner
+  attribute is a CSS property and no utility pins it. Every mark names an explicit
+  stroke width, an explicit stroke and an explicit fill, because a shape left to
+  its defaults either paints opaque black over its own labels or resolves its edge
+  to the surface it sits on. A relation naming a node it does not have is dropped,
+  and the count of dropped relations is on the element rather than swallowed.
+  
+  `ProductMark` promotes the two-part mark the product row and the product switcher
+  each hand-wrote into one catalogue item. The ring is `brand-ink` and the core is
+  `primary`, which is not a preference: a pastel brand value is a fill and never
+  text, so the core may be `primary` and the name may not, and a wordmark reads
+  `brand-ink` rather than `foreground` or `primary-foreground`. A hand-written copy
+  gets one of those two wrong and nothing in the rendered result says which. The
+  mark carries its own `data-pack` boundary, so a product's hue follows from the
+  pack the caller passed rather than from whichever pack the page is wearing, and
+  the boundary sits on an element with no radius utility, which is the placement
+  the pack-boundary law allows. A product with no pack of its own is drawn as the
+  full spectrum rather than as a colourless mark, built from the five series tokens
+  because they are the only five-way colour set the contract publishes.
+  
+  `scripts/check-vector-ink.mjs` holds both Components to the emitted contract. It
+  fails a literal colour, a gradient whose stops are not contract references, a
+  `var(--x)` naming a property the contract does not publish, and a paint utility
+  whose suffix is not a contract role, which is the last of the four a colour
+  regex cannot see because a Tailwind class is a name and not a value. Its scanned
+  set is a declared list of two file names rather than a pattern, a Component that
+  is not on disk fails the run rather than emptying it, and every exclusion carries
+  a reason and is printed on every run, with an exclusion that resolves to nothing
+  a finding rather than a line that quietly stops appearing.
+- 4f9ce77: A server-rendered pack boundary can now be mode-correct
+  
+  The only dark selector for a pack was `[data-pack="<id>"].dark`, which requires
+  the mode class on the same element. A server cannot know the reader's mode, so a
+  subtree carrying a second pack rendered in that pack's light values on a dark
+  page. The token build now emits the two-member list
+  
+  ```
+  [data-pack="<id>"].dark, .dark [data-pack="<id>"]
+  ```
+  
+  as one rule with one declaration block, for all five packs. The compound half is
+  retained, because it is what a dark surface on a light page is and it is a
+  capability the package already published.
+  
+  Consumers get a second change alongside it. `themeAttributes` now takes `mode` as
+  optional, because a required one is that same impossibility written into the type:
+  the only way to ask for a boundary with no mode class of its own was to pass
+  `'light'` as a guess, and the guess is wrong for half of readers. Omitting `mode`
+  is the honest spelling of "wear the mode of the element carrying `.dark`". The
+  change is a widening, so every existing caller still compiles.
+  
+  The theme-boundary law in `DESIGN.md` is rewritten to match. It previously
+  instructed an implementer to put `data-pack` and `class="dark"` on a boundary
+  together, which works for a client-applied boundary and cannot work for a
+  server-rendered one.
+- c126ee3: A stored theme is written when a reader decides, and not before
+  
+  `PrismProvider` wrote `prism-theme` on every resolved change, including the
+  first resolution from the server-rendered attributes or from its own defaults. A
+  first-time visitor therefore ended the load with a stored value they never chose,
+  and that value outranked the site's default on every later visit, so a site that
+  changed its default could not reach anyone who had ever loaded the page.
+  
+  The write is now gated on a decision. It happens when a caller changes the pack,
+  the mode or the toggle, and when the library recovers a decision from
+  `prism-theme-mode`, a key earlier lines wrote and this one does not. It does not
+  happen on a first resolution from the document or the defaults, so a first visit
+  ends with an empty store and a returning visit keeps its choice.
+  
+  Two readers of one stored value now share one rule. `resolveTheme` in
+  `@nanisoft/prism-ui/theming` is the rule, `PrismProvider` calls it, and the boot
+  script implements it as a string. The script cannot import the function, because
+  the build minifies and a renamed identifier inside its own `try` would throw
+  there and fail open, so the two are held together by a build-failing equivalence
+  gate that runs the emitted string and the rule over one table of stored values:
+  nothing stored, a valid pair, a valid mode with a retired pack, a valid pack with
+  a retired mode, garbage, a JSON array, a JSON string, `null`, an empty object, a
+  partial, a server-rendered pack and mode, and the retired key. Every fall-through
+  is asserted to be whole, which is the divergence this fixes: the old string
+  validated the pack and the mode independently, so a stored `{ mode: 'dark' }` with
+  no pack half-applied there and fell through whole in the rule.
+  
+  A value that is present but is not a pair is no longer repaired field by field,
+  and it is not cleared. It is the only record that this reader ever chose anything.
+  
+  The boot script records where the theme came from in `data-theme-origin` on
+  `<html>`, and it is the only writer: the provider never touches it, because two
+  writers would make the origin a race rather than a record. The value is one of
+  
+  | Value | Meaning |
+  | --- | --- |
+  | `stored` | the key held a valid `{ pack, mode }` pair |
+  | `legacy` | the key was empty and `prism-theme-mode` held a recoverable mode |
+  | `unparsed` | the key held a value that is not a pair, and the theme fell through whole |
+  | `document` | the key was empty and the root element's own attributes supplied the theme |
+  | `default` | neither did, so the consumer's defaults apply |
+  
+  It is written on every path the script reaches, so an absent origin is never an
+  unread one. It is not `data-theme`, which the migration guide retires; it records
+  the provenance of a resolution and never a pack.
+  
+  The migration clause reads `prism-theme-mode`, puts the recovered decision in
+  `prism-theme` and removes the key it read, in one guarded step. That is how it
+  expires: it fires only while a retired key still holds a mode, at most once per
+  reader, and it is retired for good by deleting the last entry of
+  `LEGACY_MODE_STORAGE_KEYS`. There is no date and no version constant in it.
+  
+  The pre-paint string gets its own ceiling in raw bytes,
+  `check-boot-budget.mjs`, because it runs before paint on every page and adding it
+  to the per-component client budget would corrupt the unit that table measures. The
+  ceiling is derived by addition: the pinned non-migration base, plus the measured
+  clause, plus one migration generation.
+- 76c8d96: Fourteen items, four catalogue gates, and a Block that publishes its interface
+  
+  **Fourteen items ship.** Three Components (`Prose`, `FactList`, `ProductSwitcher`),
+  nine Blocks (`ProcessRail01`, `StatusLedger01`, `ProductGrid01`, `StackGrid01`,
+  `LogoStrip01`, `NoteGrid01`, `InstrumentPanel01`, `SiteHeader`, `SiteFooter`) and two
+  Pages (`NotFoundPage`, `BlogPostPage`). Every one of them earns its place by the
+  growth rule rather than by taste: a marketing Block when at least three of the four
+  sites compose that section today, a Component when it has one job no existing
+  Component can express. The evidence is in the ticket and in each item's own
+  JSDoc, which names the sites and the line numbers.
+  
+  **The three contested shapes are settled as specified, and the specifications
+  held.** A process rail admits four steps and refuses five, in the type: `steps` is
+  a tuple union, so a five-element array is a compile error and not a fifth column. A
+  status ledger has four tiers rather than five, and the words for a tier are the
+  caller's: the four sites between them use eight words for those four states
+  (`live`, `available` and `complete` are one; `approved`, `specified` and `designed`
+  are one; then `planned`; then `direction`), and the vocabulary is what `statusLabel`
+  is for. A product row's mark is two parts because a filled dot fails contrast on a
+  light card, so the mark is `ProductMark`: a hairline ring in `brand-ink` around a
+  core in the product's own `primary` fill, with the name in the brand ink beside it.
+  
+  **Four gates the catalogue's own rules depend on now exist and are wired into
+  `check` and, because they read only `src/`, into `build`.** `check-item-docs`
+  holds AGENTS.md's claim that a documentation comment is the only path from an Item
+  to its published interface. `check-block-copy` holds DESIGN.md's claim that a
+  Block ships no copy and no sample data. `check-item-category` holds CONTEXT.md's
+  closed set of seven Categories and the sentence that a Block and a Page have none.
+  `check-block-imports` holds the claim that a Block fetches nothing and a Page
+  imports no router and no data client. Each resolves its roots from its own
+  location, fails a root that resolves to nothing, prints its coverage and every
+  exclusion on every run, has no report-only mode, and is proved red by a fixture
+  before it is proved green.
+  
+  **Two of those four gates were red on the tree they were written against, and both
+  findings were real.** `check-item-docs` found seven exported functions whose
+  documentation comment was not attached to the declaration: `CardHeader`,
+  `CardTitle`, `CardDescription`, `CardContent`, `CardFooter`, `SectionHeading` and
+  `Hero01`. A comment three functions above a declaration does not reach the emitted
+  `.d.ts`, so those seven were absent from the corpus. `check-block-copy` found
+  `pricing-01` rendering the word "Popular" on a featured plan that passed no badge:
+  a hardcoded claim every consumer of the Block inherited, published by the corpus as
+  part of the Block's own documentation. The fallback is gone; a featured plan with
+  no badge now renders no badge.
+  
+  **Five additive fields on four shipped items**, each with three of the four
+  consumers passing a value the item could not take, counted per consumer with a
+  file and a line and never per string. `SectionHeading.index`, for the mono ordinal
+  all four sites render above every section head (www `Landing.tsx:49`, atlas
+  `landing.tsx:48`, nexus `Landing.tsx:38`, alphalens `Landing.tsx:66`).
+  `Cta01.secondaryAction`, for the pair of actions all four closing bands render
+  (www `Landing.tsx:186`, atlas `landing.tsx:243`, nexus `Landing.tsx:221`,
+  alphalens `Landing.tsx:285`). `Cta01.note`, for the footnote three of the four
+  render under it (atlas `landing.tsx:250`, nexus `Landing.tsx:228`, alphalens
+  `Landing.tsx:292`). `FeatureGrid01.numbered`, for the ordinal all four render on
+  every card (www `Landing.tsx:166`, atlas `landing.tsx:116`, nexus `Landing.tsx:132`,
+  alphalens `Landing.tsx:142`). `Hero01.align`, for the left-aligned hero three of
+  the four compose (atlas `landing.tsx:69`, nexus `Landing.tsx:63`, alphalens
+  `Landing.tsx:87`). All five are additive: nothing changed shape and nothing was
+  removed.
+  
+  **A Block and a Page publish both an interface section and a composition section.**
+  `packages/llms` read the extractor's output for a Block and threw the interface half
+  away, so the tool told an agent a Block had no own props. The extractor reads a
+  named props type off an emitted declaration whether that declaration belongs to a
+  Component or to a Block, so both sections are now built and both are assembled into
+  the mirror file, `llms-full.txt` and the Store's `doc`.
+  
+  The Store now carries both fields for every kind, and `get_item_props` prints
+  both when it has both. The sentence "a Block has no own props" is gone from the
+  tool as well as from the published document. It was the one place the two
+  sections were still exclusive, because that tool branched on `item.props` and
+  fell back to the composition section, so a Block was answered with its
+  composition and a denial of the interface it publishes. The test asserts the
+  sentence's absence, so a refactor that restores it fails even if the composition
+  is still printed, and a second test covers the one-present case so the
+  both-present branch cannot swallow it.
+  
+  **Four accessible names became required props** on `SiteHeader`, `LogoStrip01`,
+  `NotFoundPage` and `BlogPostPage`, and a Block that ships no copy ships no
+  reader-facing copy either. `navLabel`, `productsLabel`, `label`, `linksLabel` and
+  `trailLabel` are words a screen reader reads out, and no Block should be choosing
+  them for a consumer whose site calls its own landmarks something else. This is the
+  one breaking change in the release, and it is on four new Blocks and two new Pages
+  plus one new required prop on `SiteHeader`, all of which are new in the same
+  release, so no existing call site changes.
+  
+  **`ProductSwitcher.label` has no default** for the same reason: a consumer whose set
+  is not a set of products should pass its own word rather than accept a claim about
+  what the set is.
+
+### Patch Changes
+
+- 8d24f32: A pack boundary's effect on corner radius is stated, and gated
+  
+  A scoped `data-pack` boundary resolves the pack's colour AND the pack's corner
+  radius beneath it, because radius is the only non-colour member of a pack block
+  and the token build emits it the same way. The radius range across the five packs
+  is 0.5rem to 1rem, so a row of elements each carrying its own pack shows a spread
+  of corner radii, and a small pill becomes a different pill five times over. The
+  governing law stated the colour half and was silent on the shape half, so an
+  implementer following it exactly produced mismatched corners and called it
+  correct.
+  
+  The law now states both halves and where a boundary may sit. The agent-facing
+  surface no longer says the attribute belongs on the root element, which it did
+  while listing the radius among the values it will pick up: a tool that shows the
+  value and denies the subtree is worse than a tool that is silent.
+  
+  `check-pack-boundary.mjs` holds the law. It reads the emitted `--radius-*`
+  bindings rather than a list of class names, which is the part that matters: a
+  scale step like `rounded-xl` is `calc(var(--radius) * 1.4)`, so an element
+  carrying one is pinned to a step and the step is pack-relative. The first version
+  of the gate treated "carries a radius utility" as the safe case and passed
+  exactly the defect the ticket is about. The gate also reads the colour contract
+  from the token source, so a second per-pack axis the token build has not been
+  asked about is a finding rather than a property silently counted as a colour.
+  
+  One live boundary was found and repaired: the themes page scoped five cards, one
+  per pack, each with `rounded-xl`, so the page a reader uses to compare packs was
+  comparing shape as well as colour. The cards now carry a literal radius and the
+  pack's radius is shown in the row's own label, where it is information rather
+  than an accident.
+- feb1384: The catalogue is a set comparison between three lists, not a count
+  
+  A module on disk with no catalogue entry left every count correct. It shipped,
+  `shadcn add` installed it, and `registry.json` listed it, so it was absent from
+  the corpus, the documentation site and every agent-facing tool with a fully
+  green gate run. The two lists a reader trusts, the catalogue and the registry,
+  were on the wrong side of that hole, and the checks that believed they covered it
+  compared lengths or compared a file against something derived from itself.
+  `validate-registry.mjs` holds the registry against the disk and the manifest and
+  never reads the catalogue. `sync-registry.mjs` reads the disk, so it cannot
+  disagree with the disk about what is on it. The site's
+  `test/catalogue-order.test.ts` holds the catalogue against the generated
+  ordering and the documentation tree, and both are built from the catalogue, so a
+  disagreement there is impossible by construction.
+  
+  `scripts/check-catalogue.mjs` reads the three sets independently and compares
+  them in both directions, name for name, with one printed line per offending
+  item. A length assertion cannot report which item is wrong, so nothing in the
+  run reports a length on its own: the three set sizes print as supplementary
+  coverage underneath the findings that carry the detail. All six directions a
+  three-set comparison implies are implemented, plus the kind of an item in each
+  pair of sets, the three naming rules that reconcile the three conventions, and a
+  claim that a canonical key is held by only one entry in each set.
+  
+  The registry generator deliberately keeps reading the disk. Deriving the
+  registry from the catalogue is the obvious way to remove the second list, and it
+  would delete the disagreement the check exists to see, because the generator
+  would then agree with the catalogue by construction.
+  
+  The version and the roster are two claims on two labelled lines, so a version
+  bump and a roster error cannot produce the same output. The run states its
+  coverage every time: roots resolved, files walked, files actually read, and how
+  many comparisons passed. Roots resolve from the script's own location and never
+  from the working directory, and a root that resolves to nothing fails the run
+  naming both causes.
+  
+  `PRISM_CATALOGUE_ROOT` points a run at another package root. It changes which
+  directory is read and nothing else, which is what lets the tests run the real
+  command over a fixture tree and read the real output. The gate fails on a
+  catalogue it cannot read, including an empty one: a roster that cannot be read
+  and a roster with nothing in it are different facts, and only one of them is a
+  green run.
+- 4744cce: The Slider thumb rings at full strength, and the surface check now reads both ways
+  
+  Two changes, and only two, in the component package.
+  
+  **The focus indicator.** The Slider's thumb suppressed the browser outline and
+  drew its ring at `ring-ring/50`, so the element a keyboard user lands on carried
+  an indicator that composites to between 1.14:1 and 2.74:1 against every surface
+  in all six themes. That fails WCAG 1.4.11 on the one indicator a keyboard user
+  has, and the Slider's own JSDoc says it is reachable by Tab. Thirteen other
+  components already drew that ring at full strength; the Slider was the one that
+  was missed. Nothing said so, because the contrast gate measures the `ring`
+  TOKEN and the alpha is applied in the component rather than in the token, so it
+  reads the Slider's ring as compliant.
+  
+  `packages/ui/scripts/check-focus-indicators.mjs` is the new gate. It reads the
+  class strings on the elements of every shipped Component, derives the set of
+  components whose own JSDoc claims keyboard reach or operation, and fails any of
+  them that suppresses a focus style without declaring a `ring-ring` colour at
+  full alpha alongside a `focus-visible:ring-N` width. The claim is read from the
+  JSDoc because `AGENTS.md` names that as this repository's documentation source,
+  so a component that stops claiming to be focusable is a documentation change and
+  shows up in the table the run prints rather than in a list someone maintains
+  beside the code. The unit is a class string on one element, so a Dialog's panel
+  or a Tooltip's trigger can suppress the outline legitimately; every such
+  exclusion is declared with its reason, printed with the slots it resolved to,
+  and fails the run if it stops matching anything. The run fails if the table
+  empties, and states how much it read, classified, tabled and excluded.
+  
+  **The surface check.** `scripts/check-surface.mjs` compared the published
+  surface against the source in one direction for two of its three rules. A
+  wildcard subpath that resolved to zero declarations used to leave those two
+  rules reading no file at all and the run still printed its success line, which is
+  how a published surface drifts from its source without a consumer failing first.
+  A wildcard target that matches nothing is now a finding naming the target. The
+  internal side, which the gate deliberately did not check, is now a declared
+  boundary checked in both directions: a new file under `dist/lib` is a finding
+  naming it, and a declared internal declaration that is no longer emitted is a
+  finding naming it too. The boundary and the per-target match counts are printed
+  on every run. The three existing rules and their messages are unchanged.
+  
+  **The version the four downstream site plans pin is `0.6.0`.** It is
+  `@nanisoft/prism-ui@0.6.0`, and `@nanisoft/prism-tokens@0.6.0` with it, because
+  the two are linked in `.changeset/config.json` and the pending
+  `a-server-rendered-pack-boundary-can-be-mode-correct` changeset declares a minor
+  for both from `0.5.1`. The patch declared here does not raise it: changesets
+  takes the highest bump in the queue, and a minor is already queued. A site plan
+  that pins `0.6.0` therefore resolves every subpath the export map publishes,
+  including `./components`, `./blocks`, `./pages`, `./provider`, `./theming`,
+  `./catalog` and `./styles.css`.
+  
+  No public prop, no type and no export changed. The one rendering change is a
+  ring alpha on the Slider thumb, in both modes.
+- Updated dependencies [4f9ce77]
+- Updated dependencies [5dcb347]
+  - @nanisoft/prism-tokens@0.6.0
+
+## 0.5.1
+
+### Patch Changes
+
+- 2fc6a0a: Retune `muted-foreground` so muted text clears 4.5:1 on the muted surface
+  
+  The base pack's `muted-foreground` moves from neutral 500 to neutral 600. It
+  cleared 4.5:1 on the page background but fell to 4.34:1 on `muted`, where the
+  pill, avatar fallback, kbd and tab-list pattern place it. The contrast gate now
+  also checks `muted-foreground` on `muted`. The pastel packs were already pinned
+  to their neutral 700 and are unchanged.
+- Updated dependencies [2fc6a0a]
+  - @nanisoft/prism-tokens@0.5.1
+
+## 0.5.0
+
+Prism was rebuilt from a fresh repository on a clean break. The package is now a
+published React library a consumer installs and composes, not component source to
+copy. It ships one precompiled `styles.css`, the optional `PrismProvider` and the
+`data-pack` plus `dark` theme axes, and the Components, Blocks and Pages
+taxonomy. There is no override path.
+
+Changesets are appended above this entry for every published change.

package/dist/md/components/accordion.md

@@ -0,0 +1,108 @@
+# Accordion
+
+The stack of headers that each reveal a panel in place.
+
+```tsx
+import { Accordion, AccordionItem, AccordionTrigger, AccordionContent } from '@nanisoft/prism-ui/components/accordion'
+```
+
+## Overview
+
+An Accordion is a stack of headers that each reveal a panel in place. It ships
+as `Accordion`, `AccordionItem`, `AccordionTrigger` and `AccordionContent`, all
+from the same module. Only one item is open at a time unless `multiple` is set.
+
+Each trigger is a real button, and the library ties it to its panel, so the
+header announces whether the item is expanded. Enter or Space opens the focused
+item, and the header text stays visible when the panel is closed.
+
+## Guidelines
+
+### When to use
+
+Use an Accordion when a page holds several long sections that a reader opens
+selectively: frequently asked questions, advanced settings, or a policy split
+into questions.
+
+### When not to use
+
+Do not use an Accordion when the reader needs two sections side by side. Only
+one is open at a time by default, so comparison is impossible.
+
+Do not use it to hide the primary content of a page. If a section is essential,
+it should be open, not behind a click.
+
+Do not use an Accordion for a short list of statements. A definition list reads
+better and needs no interaction.
+
+### Best practices
+
+- Give every `AccordionItem` a stable `value`, and control the open set with
+  `value` and `onValueChange` when the state matters beyond the component.
+- Open the item a reader most likely wants first, or none at all. Opening all of
+  them restores the wall of text the accordion was meant to break up.
+- Use one heading level for the triggers. Each header is a heading element, so
+  the accordion participates in the page outline.
+- Keep the trigger text a short question or label the reader can scan without
+  opening the panel.
+- Set `multiple` only when the sections are truly independent and comparing them
+  is a normal task.
+
+### Content guidelines
+
+- Phrase the trigger as the question or the topic, not as "Click to expand".
+- Use sentence case and a question mark on a question, no period on a label.
+- Keep the panel prose self-contained. A reader who opens one item should not
+  need another to make sense of it.
+
+**An accordion with one panel open and the rest available on demand.**
+
+```tsx
+'use client'
+
+import {
+  Accordion,
+  AccordionContent,
+  AccordionItem,
+  AccordionTrigger,
+} from '@nanisoft/prism-ui/components/accordion'
+
+/** An accordion with one panel open and the rest available on demand. */
+export default function AccordionDemo() {
+  return (
+    <Accordion defaultValue={['shipping']} className="max-w-measure-narrow">
+      <AccordionItem value="shipping">
+        <AccordionTrigger>How long does shipping take?</AccordionTrigger>
+        <AccordionContent>
+          Orders leave the warehouse within one working day and arrive in two to
+          five, depending on the region.
+        </AccordionContent>
+      </AccordionItem>
+      <AccordionItem value="returns">
+        <AccordionTrigger>What is the return window?</AccordionTrigger>
+        <AccordionContent>
+          Return anything unused within thirty days for a full refund.
+        </AccordionContent>
+      </AccordionItem>
+      <AccordionItem value="support">
+        <AccordionTrigger>How do I reach support?</AccordionTrigger>
+        <AccordionContent>
+          Reply to any order email, or write to support and include the order
+          number.
+        </AccordionContent>
+      </AccordionItem>
+    </Accordion>
+  )
+}
+```
+
+## Props
+
+**`value`** `string[]` · optional · default: — — The open item values, when the accordion is controlled.
+**`defaultValue`** `string[]` · optional · default: — — The item values open initially, for an uncontrolled accordion.
+**`onValueChange`** `(value: string[]) => void` · optional · default: — — Called when the open set changes.
+**`multiple`** `boolean` · optional · default: — — Whether more than one item may be open at once.
+**`disabled`** `boolean` · optional · default: — — Whether every item ignores user interaction.
+**`hiddenUntilFound`** `boolean` · optional · default: — — Whether the browser's page search may expand a panel.
+**`keepMounted`** `boolean` · optional · default: — — Whether a closed panel stays in the DOM.
+**`onOpenChange`** `(open: boolean) => void` · optional · default: — — Called when this item opens or closes.