v2.33.0

Minor — this release is about being told what you actually have. wirekit:doctor names every
personalized class block your application owns, wirekit:icons --audit separates the icon names
under contract from the ones that merely work today, the MCP server hands an assistant real worked
examples and the full component signature instead of a name and a category, and a scroll area fades
only the edges its content really continues past.

Added

• wirekit:icons --audit tells you which of YOUR icon
  names are under contract. <x-wirekit::icon> renders any name your icon set knows, so a glyph
  name works — and looks exactly like a declared alias right up until somebody switches preset, at
  which point every one of them breaks at once. The audit reads your views and separates the two,
  with the file and line of each fall-through. It never calls a glyph name an error (some glyphs
  have no alias and never will) and never suggests a replacement (checked across ten such pairs,
  ten pointed at a different character). Names bound at runtime are counted separately rather than
  quietly dropped, and a run that finds no icon usage at all exits non-zero — "nothing was measured"
  and "nothing is wrong" are different answers.

• The MCP server gained get_component_examples — worked examples instead of assembled markup.
  php artisan wirekit:mcp-serve could tell an editor what
  props a component accepts; it could not show one being used. A prop list says what is allowed, and
  an assistant filling one in guesses at the composition — which sub-component wraps which, which
  props are set together, what the canonical shape actually is. The new tool answers that from real,
  reviewed usage: 439 examples covering every documented component, and a sub-component (card.body)
  resolves to the page where it is shown inside its parent. Ask for it before writing markup.
  Documented in the AI tooling guide.

• get_component now answers with the whole component, not a summary of it. The MCP server
  described a component as a name, a category, a description and three fields per prop. An assistant
  reading that could not tell an enum from free text, saw config('wirekit.components.button.intent', 'primary') as the default and had to guess what may be passed, and — worse — was never told that
  card.body exists, which is the one composition rule the shipped guidance spends a paragraph on.
  It now returns the same picture the JSON manifest carries: the documentation URL, whether the
  component is anonymous or class-based, every declared slot with whether it is required, every
  sub-component with its own props, and the full prop signature including type hints, the resolved
  default behind a config(...) call, and the example values the docblock names. A test compares
  the two surfaces field by field, in both directions, so one can no longer learn something the
  other does not. Documented in the AI tooling guide.

• Three registry helpers became public API, because the manifest and the MCP server were each
  deriving the same answers privately and had begun to disagree: ComponentRegistry::slotsOf(…)
  returns a component's declared slots with their required flag, ComponentRegistry::describeSubComponentsOf(…)
  returns its sub-components with their props, and ComponentRegistry::existingBladeFilePath(…)
  resolves a component's template — or null when it has none, which is the answer the
  path-returning companion cannot give.

  The examples are extracted when the package is built rather than read at runtime, because the
  documentation is not part of what gets installed — a server that read it would answer correctly in
  WireKit's own repository and "no examples" in yours.

• scroll-area gained fade="auto" — an edge
  fade that measures before it masks. The named edges (both, start, end) are unconditional
  CSS, which is what makes them free and also what limits them: they fade the top edge while the
  reader is already at the top, the bottom edge at the bottom, and both edges on an area whose
  content fits and cannot scroll at all — taking ink off text that is entirely visible, to signal
  something that is not true. auto masks only an edge the content continues past, and follows
  content that arrives later, which is the case worth having it for: a transcript appending a
  message, a list a search filters down, a panel that opens. Nothing to call, nothing to refresh.
  It is the one value that needs JavaScript, and it fails toward no mask at all rather than the
  wrong one — a missing hint instead of dissolved text. The named edges are untouched and stay pure
  CSS, so nothing that exists today changes, and the depth is still the --fade-wk-size token.

• wirekit:doctor names every personalized block that
  replaces the shipped one.
  WireKit::personalize() takes two value shapes per
  block, and they differ in a consequence nothing reported: a finished class string REPLACES the
  block, while a closure receiving the vendor default extends it. A replacement is a valid choice —
  it also ends the flow of later WireKit changes to that block, permanently and without a word, so
  the personalization keeps looking like a decision somebody made long after it has stopped
  inheriting improvements. The check reports replacements as a warning with the block names and
  offers the closure form for the case where only a delta was wanted. It stays silent when every
  block extends. A new WireKit::personalizedComponents() returns the names of the personalized
  components; the map could be read per component but never enumerated, and a diagnostic cannot
  guess names it has no way to list.

• wirekit:doctor:props gained --require-in-scope.
  A run that scans real templates and finds none of them using a WireKit component has two honest
  readings, and which one is right depends on the application rather than on the linter. If you do
  not use WireKit in that tree, nothing in scope is correct and the default still succeeds. If you
  use it everywhere, the same result means the walk found the wrong tree — a second view path, a
  renamed directory, an argument pointing somewhere empty — and a green run is the last thing you
  want. The flag is how you say which application you are. Reported by a developer whose only handle
  on that state was matching the success sentence in a shell script, which a reword would have
  deleted silently.

Fixed

• The tab bar's active indicator was missing from the compiled CSS. A tab bar's appearance
  moved into PHP in 2.31.0 — a good refactor, and one Tailwind's @source glob never looks at, so
  six classes stopped compiling. They were not decoration: they are the active-tab indicator and
  the margins that pull it onto the container edge. The bar rendered, the tabs worked, the ARIA was
  correct, and the selected tab was simply not marked. Reported from a project that attributes its
  built stylesheet byte for byte — 123702 → 123169 bytes, six selectors gone and none added. Fixed
  through the safelist mechanism that already exists for this class of bug, listing all 48 emitted
  classes rather than only the six that went missing: the other 42 survive today because some
  unrelated view happens to use flex or gap-1, which is a coincidence and not a guarantee.

• <x-wirekit::fonts> overwrote the shipped font
  tokens with weaker ones. The component wrote all three --font-wk-* variables unconditionally,
  standing in a hardcoded stack for a category nobody had configured — and those stand-ins are
  shorter than what the package ships. Both declarations sit unlayered at equal specificity, so
  document order decided it, and placed after @wirekitStyles the monospace stack silently lost two
  families. Nothing threw and the markup was identical either way; it showed only to a reader who
  had those fonts installed. An unconfigured sans or mono is now simply not declared, so the
  stylesheet's value stands from any position. Serif is deliberately still written, because the
  stylesheet does not declare it and omitting it would drop every serif surface to the browser
  default.

• The fonts.fallbacks example gave a real family another font's numbers. The configuration
  stub and the fonts page both showed a named family
  with measurements that belong to a different one — directly below a line reading "measure the four
  values, do not estimate them". A developer whose font really was that family read the block as
  already measured and pasted it. Both examples now use a placeholder family with blank
  placeholders: a blank cannot be copied, a plausible number can.

• wirekit:doctor told developers to delete configuration that was working. The config-drift
  check compared key names against the shipped stub and reported anything the stub does not carry as
  an option "this version no longer offers". Reported from a project where ten keys were named and
  all ten were wrong, in three shapes — keys whose names belong to the developer rather than to the
  stub, a feature whose stub value is an empty array so no correct use could ever match, and leaves
  sitting under a branch the stub does carry. Developer-keyed nodes are now exempt, a path that is a
  prefix of a stub key is not an orphan, and the wording no longer asserts that nothing reads them:
  a diff is evidence, not a verdict. The list also prints in full, because the truncation hid half
  of a finding whose whole point was which keys were named.

• An underscore-spelled regional locale resolved to the wrong variety. pt_BR and pt-BR are
  one locale wearing two separators, but only the base-language half of that was handled. The
  underscore spelling never reached the regional catalog this package ships: it missed, the base
  fallback answered, and a Brazilian-Portuguese application quietly rendered European Portuguese.
  Nothing threw — the strings were all present, just from the wrong catalog.

• The component manifest said nothing about glass. Of the whole catalog, exactly one component
  carries no props, no slots and no sub-components, and its emptiness is real. In a manifest that is
  indistinguishable from a component whose props could not be parsed — and the wrong reading is the
  expensive one, because a tool that assumes a parse failure will invent an API. Its description now
  says so, and a guard requires any component with nothing to declare to declare that.

• list reported that it accepts no content. Every machine-readable surface — the JSON
  manifest, the project-root schema file, and every tool fed by them — listed the component with an
  empty slot array while its template renders {{ $slot }} on its last line. A developer asking
  the manifest how to use it was told to write an empty tag. The cause was a second Blade-path
  resolver that knew the flat and dotted filenames but not the directory-index form list is
  written in, so the file was never found and "no template" read as "no slots" — indistinguishable
  from the components that genuinely have none. Both surfaces resolve through one path now, and a
  test fails whenever a component that renders a default slot fails to report one.

• Two siblings of the same overlay no longer stand open at once. Opening a second
  popover,
  dropdown,
  hover-card,
  menubar or
  navigation-menu menu on the same page left
  the first one open behind it, and the two panels overlapped. Nothing reported it — no console
  error, no changed markup — because the only symptom is what a reader sees. context-menu and
  combobox had each solved it separately; all seven now share one mechanism, so the next overlay
  inherits it instead of copying it. Opening a dropdown still does not close a popover: the
  coordination is per component family, which is the behavior that existed before and is not a
  question a patch release should answer differently.

• The CSP advice about method names was far wider than the rule it described, and the extra
  width cost real renames. Two pages and
  wirekit:csp-audit itself said that a Livewire method
  "whose name is a JavaScript keyword" needs index access — $wire['delete'](...) instead of
  $wire.delete(...). Measured against the parser that decides it, that is true of ten names and
  false of the forty-two other reserved and future-reserved words, because a reserved word after a
  dot has been an ordinary property name since ES5. A developer auditing their own component
  against the old sentence renames public action names — for, class, return — that were never
  affected, and every one of those renames is reachable from templates and tests. All three places
  now print the set itself: delete, false, in, instanceof, new, null, true, typeof,
  undefined and void, with a named counter-example so the list reads as complete rather than as
  a sample. The set is no longer written by hand anywhere — it is checked against the tokenizer's
  own table on every test run, so a future change to that table fails the build instead of leaving
  three pages quietly wrong.

Documentation

• Two PHP entry points that were only findable in the source are now on a page.
  WireKit::avatarPaletteFor(...) returns the same background/foreground pair
  <x-wirekit::avatar from-initials> derives, so a custom chip can match an avatar without rendering
  one. WireKit::defaultsFor(...) reads back what WireKit::defaults([...]) registered — with the
  distinction stated on the page, because it reports the runtime record rather than the value in
  effect.

• Four tables across three pages rendered as raw text and now render as tables. A blank line, a
  callout and a paragraph had each been placed inside a table, and one font-size table carried no
  header row at all. Markdown ends a table at the first interruption rather than resuming it
  afterwards, so every row below the break was published as pipe-separated body text. Affected the
  bundles table on the dependencies page, the App Shell prop table, and the motion and font-size
  token tables.

• wirekit:verify --tier examples corrected. The commented check ranges beside the two --tier
  examples still described the numbering from before a check was added, and the environment-tier
  comment pointed at a package-tier check.

• The gap scale is now reachable from the components that use it. --gap-wk-* and
  --space-wk-* do not run on the same ladder, and the
  design-token page has said so since 2.28 with
  three guards keeping the table honest. It was still reported twice from two applications eight
  days apart, and the second report was measured against a version that already carried the table.
  That is a placement problem rather than a documentation one: measured, the words gap-wk,
  design-tokens, ladder and rung appeared zero times on the row, stack and grid pages, so a
  developer typing gap="lg" had no path to the paragraph that prevents the mistake — which also
  had no heading, so it was neither linkable nor in the page's contents. It has one now, and the
  five components that let a developer name a rung link to it:
  row,
  stack,
  grid,
  bento-grid and
  feature-grid.

• The customization page no longer says nothing warns
  you. Its section on adjusting a block rather than replacing it explained that taking ownership
  silently stops later improvements from reaching it, and closed with "nothing warns you, because
  nothing is broken". The doctor check in this release is that warning, so the sentence became
  false in the same release that made it obsolete. It now names the check.

• The PHP discovery surface is documented where it claims to be. The
  ComponentRegistry page opens by calling
  itself the canonical surface for discovering every component, and documented five of its thirteen
  entry points — subComponentsOf, tag, tagAlias, resolve, componentClass,
  isSubComponent, subComponents and extractAwareProps were all shipped and described nowhere.
  All of them are there now, with the three this release adds. Two claims on the page were also
  wrong: type_hint was said to be always null where fifteen props carry one, and the prop-record
  table listed five fields where the record has six.

• The theme-preset registry had only its write side on a page.
  ThemePresetRegistry::register() was documented; the
  four reads that make it useful were not — which is exactly the set a theme picker needs.
  all(), keys(), get() and isValid() are documented together now, with isDefault() called
  out separately because it is the one a picker gets wrong: default is not a preset with variables,
  it is the instruction to remove the block, and treating it as a normal preset writes an empty one.

• WireKit::cspNonce() and WireKit::prefix() are documented — the first in the
  integration guide's nonce section, with
  its resolution order and the fact that null means "no policy" rather than a failed lookup; the
  second in Getting Started, where code that builds a
  tag name should ask rather than assume, because the prefix is a setting and wirekit is only its
  default.

• Localization no longer reads as base-language only.
  The page described the shipped catalogs without saying that a regional variety is a catalog of
  its own, which is the half a developer needs before choosing a locale string.