# Changelog ### 4.12.0 - 2026-08-08 #### Added - **`combo_box` M4 - field citizenship and sizes.** `<.field type="combobox">` brings the full field grammar - label, changeset errors (with the wrapper error ring on the control and trigger), `help_text`, required - and forwards every combobox attr; `combo_variant` selects the anatomy (`:variant` belongs to radio-card), the field size family maps onto the new combobox `size` attr (`sm`/`md`/`lg`; `xs`/`xl` clamp), and `clearable` passes through. Multiple gets the empty hidden input so forms post cleanly. Rich slots (`:option`, `:selected`, `:chip`, `:header`/`:footer`) stay on the bare component by design. #### Added - **`combo_box` M3b - the modes: `free_text`, `create`, remote search.** The last of the Tom Select parity surface, on the new architecture: - `free_text`: Enter at the empty stop commits the typed text as a dynamic option in the hidden select, so forms post it like any choice. Existing labels are chosen instead of duplicated (case-insensitive). The server owns persistence. - `create`: free_text plus an explicit, keyboard-reachable "Create …" row (the last arrow stop) that appears for novel queries. `create_label` localizes the verb. - `remote_options_event_name` + `remote_options_target`: the contract preserved verbatim - typing pushes the raw term (debounced 300ms), the handler replies `{:reply, %{results: [%{text: …, value: …}]}, socket}`, and the hook renders the results (the listbox is hook-owned in remote mode - one writer per region). Designed loading row (`loading_label`), stale replies dropped by sequence, chosen results inserted into the select. #### Added - **`combo_box` M3a - the slot family completes: `:header`, `:footer`, `:selected`, `:chip`.** All four speak the `:option` slot's grammar (`:let` receives normalized options with `meta`), all render-only, no new hook state machines. `:header`/`:footer` are panel chrome OUTSIDE the listbox - captions, counts, manage links - keyboard navigation and filtering never touch them. `:selected` renders rich closed-state content in the trigger (colored label dots with a "+N" overflow is pure composition, not an attr); `:chip` renders rich chips with the remove button intact. Both closed-state slots reconcile by server-wins-on-patch: client picks show plain optimistic text, the LiveView patch swaps the rich content back in, and fresh server-rendered DOM is left untouched whenever it already matches the selection (value-stamped diffing on the label and per-chip `data-value`). - **`DataTable.State` + the in-memory engine - the 4.12 data table foundation.** `PetalComponents.DataTable.State` is the whole backend contract in one struct (multi-sort `order_by`, typed `filters`, page, page_size, total) with `from_params/2` / `to_params/2` for URL-as-state round-trips (whitelisted fields, no atom creation from user input, clamped sizes) plus the interaction helpers (`toggle_sort/2` cycling asc→desc→off, `put_filter/4`, `clear_filters/1`, `total_pages/1`). `Engine.List.run/2` sorts, filters and paginates a plain list against that state - case-insensitive strings, nils always last, numeric and date coercion from string params - so registry examples and the playground get a working table with zero setup. The `<.data_table>` component builds on this next. #### Added - **`combo_box` trigger variant supports `clearable`.** A chosen single-select trigger had no road back to the empty state (Nic's find - the input variant had the X, the trigger silently ignored the attr). A button can't nest inside the trigger button, so the clear is a sibling positioned over the trigger's right rail before the chevron (Base UI anatomy), gated by the same `data-has-value` state, with a 24px hit target. Clearing empties the select, restores the placeholder label, keeps the panel closed and returns focus to the trigger. Multiple mode stays clear-less by design on both variants - chips and panel toggles are its road back. #### Fixed - **`combo_box`: iOS taps on control chrome and the trigger button reliably toggle the panel.** iOS Safari fires `focusout` with `relatedTarget: null` even when focus moves WITHIN the component, so a chevron or trigger tap closed the panel mid-press and the tap's own click instantly reopened it - an invisible flash that read as a dead chevron zone (input variant) and a picker that never closes (trigger variant). Device-log verified. The focusout close now defers one tick and verifies where focus actually landed instead of trusting `relatedTarget`; genuine Tab-away and click-away close exactly as before. - **`combo_box`: a touch-scroll no longer dismisses the panel.** The outside dismiss closed on `pointerdown`, so starting a scroll gesture anywhere outside the panel killed it the instant the finger landed (Nic's iOS find). The dismiss is now a completed press - `pointerdown` arms it, `pointerup` in (roughly) the same spot closes; a scroll ends in `pointercancel` or a far-away release and leaves the panel open, matching shadcn/Base UI and the old Tom Select behavior. A clean tap outside still closes, and the desktop focusout path is unchanged. #### Fixed - **`combo_box`: chevron press reliably toggles the panel closed on every platform.** Two interacting bugs: on desktop, pressing the (non-focusable) chevron blurred the input, focusout closed the panel mid-press, and the click then saw a closed panel and REOPENED it - the toggle flashed instead of closing. On iOS no blur fires, but tap- target correction rewrites the synthesized click's target and coordinates onto the nearby text field, so the tap read as caret work and nothing closed. Fix: chrome-vs-caret is decided at `pointerdown` (which hit-tests the real touch point and is never rewritten), the click consumes that record, and chrome presses are preventDefaulted so the input never blurs mid-press. Input presses keep caret behavior and never discard the active query. #### Fixed - **`combo_box` multiple: the highlight stays on the item just picked.** Choosing used to reset the filter and re-home the highlight on the first option - a disorienting jump when picking from the middle of the list (Nic's on-device find). Now arrowing resumes from the item you just toggled (Base UI/downshift grammar); the trigger variant's multiple mode gets the same fix through the shared path. #### Changed - **`combo_box` multiple: the placeholder stays visible with chips present** ("Add members…" style - the reui/Mantine TagsInput convention) instead of vanishing after the first pick, and it rests while `max_items` is reached (pure CSS on `data-max-reached` - the attribute itself stays server truth, so live placeholder changes always win). Single-select behavior unchanged. #### Added - **`combo_box` `:option` slot - rich options.** Render anything inside each panel option (avatars, flags, secondary text); `:let` receives the normalized option - `label`, `value`, `disabled`, and `meta` carrying whatever extra data the option tuple's keyword list held (`{"Amelia", "am", role: "Engineering"}`). Filtering, chips and the trigger label keep using the plain label, so rich content never affects search or the closed state. - **`combo_box` `variant="trigger"` - the picker anatomy.** A select-like button shows the chosen value (or "N selected" with `multiple` - `count_label` localizes the word) and the search input lives inside the panel, command-palette style. Open from the button, arrow keys or click; search; choose - single select closes and returns focus to the button, multiple stays open and the count label tracks live. Same hidden select underneath, same form behavior, same positioning (flip + space cap). This is the anatomy pickers use - and the exact shape the 4.12 data table's filter editors are built from. - **`combo_box` grows up: `multiple` with chips, `clearable`, and the hardening pass** (combobox M2 core). `multiple` turns the trigger into a chip row - removable tokens, panel stays open while picking, Backspace in an empty input removes the last chip - backed by a real ``. The select IS the form control - name/value from a `field` or manual attrs, bubbling input/change on selection - so changesets, `phx-change` and LiveView form recovery behave exactly like a plain select. No `phx-update="ignore"` island, no Tom Select, zero JS dependencies. Options accept the `select` shapes: strings, `{label, value}`, `{label, value, disabled: true}` and `{group_label, options}` groups (headed in the listbox, `optgroup` in the select, hidden when filtered empty). ARIA done properly for a combobox: `aria-selected` marks the CHOSEN option (with the check mark); the keyboard highlight is virtual (`data-highlighted` + `aria-activedescendant`), focus never leaves the input. The control carries the field surface (input_group doctrine: ring on the wrapper, borderless inner input); the panel follows the dropdown grammar with the 1rem radius clamp. Single select, static options; multiple/chips, free-text autocomplete and the remote-search contract land in the next milestones. - **`pagination` gains `variant="simple"`** - Previous/Next outline buttons with disabled boundary states, instead of the numbered page window. The right form when a list is short or a total is unreliable (cursor-style paging), and the footer grammar the upcoming `data_table` uses when `total` is unknown. `previous_label`/`next_label` attrs make the copy localizable; link and event modes work unchanged. The numbered layout is untouched and stays the default (`variant="numbered"`). ### 4.11.3 - 2026-08-05 #### Fixed - **Adjacent pressed toggle chips no longer merge into one block.** The solid variant's pressed chip drew its definition edge as an outward `ring-1` - a box-shadow with 1px spread, which paints outside the border box without taking layout. In a multi-select rail two neighbouring pressed chips each bled 1px into the rail's 2px gap, so the gap read as zero and the chips looked welded together (outline and accent set `ring-0` and never showed it). The ring is now inset: same edge, gap restored, and every variant separates its pressed chips identically. - **The input_group composer toolbar uses square icon buttons.** Its three formatting buttons were `size="xs"` holding an icon, so they rendered as 38x30 pills while the same Bold button in the button_group and tooltip examples - `size="icon"` - is a 34x34 square. The example now uses the size built for the job. No component change: `size="icon"` and the full `icon_button` ladder (36 through 56, all square) were already there. #### Changed - **Every toggle group variant washes on hover, not just outline.** Solid and accent shifted text colour alone, which is a weak tell that a chip is clickable - and it meant one of three variants behaved differently from the other two for no reason anyone could name. The wash now lives on the base item so all three get it, translucent so it composites over whichever rail it sits on; outline keeps its own tuned step (its rail is transparent, the others sit on a wash). Pressed chips are untouched - their rules land later at equal specificity or higher, so pressed paint survives hover - and disabled items are explicitly excluded in both forms, including the radio form where `disabled` lives on the input and needs the `:has()` rule. - **Custom scheme-switch icons centre instead of huddling top-left.** The scheme wrappers force slotted children to fill the box, which is right for masked hero spans but left text content - the emoji example - pinned to the top-left corner of a block box. The forced child is now a centring flex box: text and emoji slot content centres, masked icons render identically (the mask fills the box regardless of display type). - **Icon-only toggle chips are squares, not pills.** The rail's per-size horizontal padding made icon-only items wider than tall (36x28 at md), while the scheme toggle - the component it sits next to in every toolbar - is a perfect square. Icon-only items now drop their horizontal padding to match the vertical (28x28 at md), keyed off the aria-label the icon-only contract already requires: it sits on the item itself in multiple mode and on the hidden radio in single mode, and items with visible text must not carry one (the slot attr doc now says so). The comment above the icon sizing rule always claimed "a square hit target" - now it's true of the chip, not just the icon. ### 4.11.2 - 2026-08-05 #### Fixed - **Two icon-geometry sites the 4.11.1 sweep missed join the `!important` contract.** 4.11.1 restored the bangs at the sites the 4.11.0 audit had de-banged - but two rules were never in that audit because they never had bangs to begin with, and both lose the same cascade fight (consumer hero-* geometry from the utilities layer or unlayered): the toggle group's icon-only item sizing (on petal.build the pressed chip stretched to 24x20 and the rail read non-uniform, icon riding left) and the alert kind-glyph fill inside its 20px container. Both banged, both verified with the same compiled-CSS probe against unlayered consumer hero rules: 16x16 icons, uniform 36x28 chips. ### 4.11.1 - 2026-08-05 #### Fixed - **The icon sizing contract gets its `!important` back - 4.11.0's de-bang broke icon geometry in every phx.new-style app.** 4.11.0 dropped `!important` from the nine icon-sizing sites on the belief that a consumer's heroicons plugin lands in the components layer, where the doubled selector's (0,2,0) wins on specificity. That premise is false in the two setups that matter: Tailwind v4 `@plugin` (what every phx.new app vendors, petal_pro included) emits hero-* rules into the *utilities* layer, which beats any components-layer rule regardless of specificity - so the plugin's `width: spacing.6; height: 1lh` won and avatars' placeholder silhouettes shrank off-centre while scheme-switch glyphs rode high in their boxes. And generated heroicon stylesheets often sit unlayered, which beats layered rules too. The doubled selectors stay (they carry flattened builds, where specificity does decide) and the geometry declarations are `!important` again, including the colour-scheme wrappers that had shipped bang-less since 4.9 and were quietly broken in layered builds the whole time. Verified by compiled-CSS geometry probes against a phx.new-style build: avatar placeholder back to filling its circle, scheme glyphs back to 20x20 centred. The honest cost is restored too and documented at the contract's canonical comment: these internal icons are fixed-size, and even a bang utility can't resize them. - **Toggle group radios stay invisible mid round-trip.** The hidden radio that fills each toggle_group item was hidden by `opacity: 0` alone - but app CSS routinely forces opacity back up: phx.new's default `phx-click-loading` styling is a *utility* at opacity .5, and utilities beat the components layer. So with `@tailwindcss/forms` loaded, every server round-trip flashed the full-chip-size radio at half opacity - blue fill, white centre dot, focus ring - for the length of the round trip plus a 1s fade. The input now paints nothing even when visible: `appearance: none`, no background, no border, no box-shadow, all in the components layer where they beat the forms plugin's base-layer styling. Reproduced and verified against the compiled playground stylesheet with the loading class applied: 4.11.0 shows the dot, patched shows a normal pressed chip. ### 4.11.0 - 2026-08-05 #### Added - **`hide_header` on `<.modal>` - headerless confirmation dialogs without an accessibility trade-off.** Confirmation-style modals (think "unsubscribed successfully") carry their heading centered in the content, but the component always rendered its header bar, so an empty strip sat above the content unless you hand-rolled CSS to hide it. `hide_header` now does it properly: the header collapses to a screen-reader-only box (sr-only, never `display: none`) because the dialog's `aria-labelledby` points at the title inside it - pass a real `title` and the dialog keeps its accessible name while showing none of the bar. It also skips the corner close button (an invisible button you can still tab to is a keyboard trap, not a feature), so give the content its own way out - escape and click-away still work by default. New "Headerless confirmation" example in the showcase registry. - **`tailwind-gray.css` - stock Tailwind gray back in one import.** The `gray` role deliberately ships zinc values under the gray name, which also decides what an app's own `text-gray-*` utilities render as - and since the stock values only ever lived under that name, no `var()` can reach them once zinc holds it. This file restores them verbatim: import it after `default.css` and every gray in the app is Tailwind's original again. Any other neutral (slate, stone, neutral) never needed it - those remain live variables you can remap a role to in one line per stop. - **The toggle group gets a live example.** Every toggle_group registry example rendered inert, so the docs pages read as screenshots. The new "Try it - this one is live" example is the exception on purpose: the single rail is native radios, so the browser moves the pressed chip with no JavaScript and no server, and the multiple rail toggles each chip client-side through `JS.toggle_attribute` on `on_change`. The description says out loud what a real app would do instead (send the event, let the server own the value). #### Changed - **The alert page leads with the semantic ramps in outline.** Same first example, same anchor, but the four alerts now carry `variant="outline"` - the coloured border and glyph on a calm surface reads better at first glance than four filled light bars. The description points out that dropping the variant gives you light, the component default, which is unchanged. #### Fixed - **A consumer `@theme` block placed before the petal_components import no longer loses silently.** The shipped colour ramps were a plain `@theme` - a hard declaration - so an app that defined its brand palette above the import (the natural place for it) got every ramp quietly replaced with our defaults, `primary` included, with nothing to say why. The ramps are now `@theme default`, the same soft-fallback convention Tailwind uses for its own palette: your `@theme` wins wherever it sits, and a fresh install with no overrides renders exactly as before. Verified against Tailwind 4.3.1's compiler in both orderings. - **Component icons can be resized again - icon sizing drops `!important` for a real contract.** Nine icon-sizing sites (pagination and menu chevrons, accordion plus/minus, the avatar placeholder silhouette, the chat family's reasoning chevron, marker, action-bar and composer-send icons) defended their geometry with `!important`, on the belief that a consumer's heroicons plugin sizes `.hero-*` in the utilities layer. It doesn't: the plugin every phx.new app vendors registers icons through `matchComponents` - the same components layer as our rules, at single-class specificity - so the fight was always a coin flip on import order, and `!important` was both overkill and a lockout (important component declarations beat even a consumer's `w-6!`, so those icons were genuinely un-resizable). All nine now use the doubled-selector contract the colour-scheme controls proved out in 4.9.0: `(0,2,0)` outranks the plugin in either import order, and a plain `w-* h-*` utility on the icon overrides cleanly via the utilities layer. The carousel's slide-mode positioning also drops three `!important`s its own selector specificity already made redundant, and the two keepers now say why inline: the showcase `
` background fights an inline style from mdex's `:html_inline` formatter (only `!important` outranks an inline style), and the reduced-motion `1ms` override is the standard accessibility pattern.
- **The rest of the bang utilities go too - the whole file is now `!important`-honest.** The legacy `has-error` form styling (label, span, text inputs, selects, textareas, the file-input rule) carried a pile of `!` suffixes it never needed: form.ex puts `has-error` on the same element as the pc class, so `input.has-error` at (0,1,1) outranks every single-class base rule, and its focus variants at (0,2,1) outrank the base focus-visible styling - specificity was always enough. Same story for the scheme dropdown's positioning overrides (`!w-auto !min-w-0 !left-1/2 !right-auto` refine base dropdown rules they already beat on source order) and the checkbox label's `mb-0!`. All de-banged with zero rendered change, each site commented with why no bang is needed - and consumer CSS can now override every one of them the normal way. The audit also unearthed one rule that was never alive: `input[type="file_input"].has-error` targeted a type attribute nothing renders (`Form.file_input` has always emitted `type="file"`), so it's removed outright - which also means error-state file inputs have never had danger styling; giving them some for real is a separate design decision.

### 4.10.1 - 2026-08-04

#### Fixed

- **Showcase frames no longer trap overlay examples under the host page's chrome.** `.pc-showcase__preview` carried a `z-10`, and a z-index on a positioned element creates a stacking context - so the modal and slide-over examples embedded on docs pages had their `fixed z-50` overlays flattened to an effective z-10 against the page. On petal.build, the docs sidebar (`fixed z-20`) and navbar painted over an open slide over, and the dark overlay skipped them. The `z-10` had no layering job (the preview and code panels are non-overlapping siblings; nothing in the stylesheet uses a negative z that relied on the context) so it's simply removed, and the rule now carries a comment explaining why it must never get one back. Raising the overlay's z-index could never have fixed this - a z-index only competes inside its own context. Verified in headless Chrome against the compiled stylesheet: with the trap, a z-20 sidebar wins even against an overlay forced to z-9999; without it, the overlay covers everything, and page content behaves identically. **Scope: docs surfaces only** (petal.build and the playground render showcase frames; consumer apps never do). The command palette was never affected - its native `` lives in the browser's top layer, which no ancestor stacking context can trap. That immunity is the direction of travel for modal and slide over.
- **Opening a modal or slide over no longer nudges the layout around it.** The root element each renders (`.pc-modal`, `.pc-slide-over`) was an in-flow div - a 0x0 box, since its overlay and panel children are `fixed` - but in-flow means it becomes a real flex/grid item the moment JS reveals it. Inside any gapped or centered container that costs a `gap`: on the petal.build docs pages the trigger button jumped 8px left on open and back on close (the showcase preview is a centered flex row), and the same happens in any consumer toolbar or button row built the same way. Both roots are now `fixed`: out of flow, zero layout contribution, opening an overlay can never move its siblings. The overlay and panel were always viewport-fixed, so nothing about how modals and slide overs look or position changes - verified in headless Chrome against the compiled stylesheet (8px shift before, 0px after, panel and overlay placement identical). The one consumer-visible nuance: custom CSS that absolutely positioned something against the root's old `position: relative` would need its own positioning context now.

### 4.10.0 - 2026-08-04

#### Added

- Nine new registry examples across six thin modules - toast's remaining
  severities (warning, loading, neutral), carousel thumbnail navigation
  and a slides_per_view row, the avatar size ladder + presence dots and
  the avatar_group team stack, marquee's vertical feed, local_time's
  threshold flip, and the colour-scheme switch labelled in German plus
  custom icon slots. petal.build renders these registries directly, so
  the docs pages inherit every example in the same release.
- `language_select/1` graduates from Petal Pro: the locale dropdown, now
  a free composition of `dropdown/1`. Plain links carry `?locale=`
  (`&`-joined when the path already has a query - the Pro version broke
  there), the current language gets `aria-current` plus a check, and an
  unknown `current_locale` falls back gracefully instead of raising.
  Beyond the port: `variant="code"` / `variant="label"` put text on the
  trigger instead of a flag, options may omit `:flag` entirely - a flag
  is a country, not a language - and `show_chevron={false}` leaves the
  trigger standing alone.
- `social_button/1` graduates from Petal Pro - the very first Pro
  component (May 2022), now free and rebuilt on `pc-button` geometry so
  provider buttons follow the theme, sizes and radius like every other
  button. Nine providers (google, github, apple, x, facebook, microsoft,
  gitlab, discord, linkedin - the 2022 Twitter bird retires), outline
  default with coloured glyphs, solid brand paint, auto "Continue with"
  labels, and `icon_only` with the label as the accessible name.
- `brand_icon/1` - the nine provider glyphs as standalone inline SVGs,
  `currentColor` by default, `colored` for the official brand treatment
  (gitlab and discord vendored from simple-icons, CC0). Heroicons ships
  no logos; now the package does. The Card login example dogfoods the
  pair.
- **`command_trigger` - the visible opener for the ⌘K dialog.** The search pill every command-palette app puts in its header (magnifier, label, kbd hint) plus an icon-only variant for tight rows, so consumers stop hand-rolling it - marketing, the playground topbar and the docs example had three bespoke copies between them. Opens the `command_dialog` named in `dialog_id` through the new `PetalCommandTrigger` hook rather than a phx-click JS command, so it works on dead views too (hooks mount there since LiveView 1.1; JS commands do not run). The shortcut itself still lives on the dialog - the kbd chip is the visible hint, and `kbd={nil}` hides it.
- **The segmented family joins the button radius curve.** `color_scheme_switch`'s segmented variant and `toggle_group` now round at token x1.2 (rail) over x1 (chips) instead of the old pill-at-default x2.4/x2. Every radius-token step now produces a distinct, perceptible corner change - the old curve spent its whole range between 0 and the default and then saturated - and only `full` pills, exactly like buttons. One curve across the system; at the shipped default the rails read rounded rather than pill. Nested menu items complete the same doctrine: the scheme dropdown's options and the command palette's items now derive their radius from the clamped panel (the dropdown-item rule) instead of the raw token, so `full` can no longer pill an item inside a 16px panel - and the scheme dropdown adopts the dropdown's hover/selected wash ramp while it's there.
- **`toggle_group` - a segmented selection rail.** Where `button_group` groups actions, `toggle_group` holds a selection: one pressed option by default, any number with `multiple`. Server-driven and stateless the LiveView way - pass `value`, handle `on_change` (the pressed option arrives in `phx-value-toggle`), no hook, no client state. Single select renders native radios (the scheme-switch segmented mechanics): real radiogroup semantics, one tab stop, arrow keys move the selection. Multiple renders `aria-pressed` toggle buttons. Three variants - `solid` (wash rail, neutral chip), `outline` (bordered toolbar rail on the outline-button ramp) and `accent` (the selection painted in the brand colour) - three sizes on the scheme-segmented radii math, per-item or whole-rail `disabled`, items take any content. The radios detach from surrounding forms, so a rail inside a `<.form>` never posts a stray param. Ships with registry examples (including the device rail) and a playground page with live demos.
- **`review_card` works without a photo.** `img` is optional now: leave it off and the card hashes a deterministic gradient monogram from the reviewer's name - testimonial walls with zero photo assets. With `img` set, nothing changes.
- **Showcase registry: marquee grows the striking spread.** The logo strip (styled wordmarks - swap in your `` logos and nothing else changes) and the testimonial wall (two counter-rotating rows of photo-less review cards). The playground marquee page renders both.

#### Changed

- Dropdown menus size to their content: the panel's fixed `w-56` becomes
  `w-max min-w-36 max-w-72`, so short menus (user menu, language select)
  stop floating in dead space while long items still cap at a sane width.
  Pin a fixed width back with `menu_items_wrapper_class` if a layout
  relied on it.
- The user menu's trigger chevron had drifted bright (`dark:text-gray-100`)
  while the rest of the dropdown family runs muted - both now share the
  new `.pc-dropdown__chevron` class, so the colour can't drift again.
  `user_dropdown_menu` also gains `show_chevron={false}` for the
  chevron-less avatar trigger.
- **LiveView floor moves to `~> 1.1`** (Matt-approved, 3 Aug). The 4.8+ components already relied on 1.1 behaviour - the colour scheme switch and command palette mount hooks on dead views, which 1.0 never did - so the requirement now states what the package actually needs. Anyone still on LiveView 1.0 stays on petal_components 4.9. This also makes `command_trigger`'s dead-view support unconditional across the supported range.

#### Fixed

- `toggle_group` items: `aria-label` is now a declared slot attr.
  Icon-only items were passing it through with a compile warning - fatal
  for consumers building with `--warnings-as-errors`.
- Tall elements no longer inflate into ellipses at `radius="full"`: the
  panel clamp (1rem ceiling) now also covers textareas, multi-selects,
  block input groups (the composer pattern), alerts, the chat composer
  and tooltips. One-line controls keep their pills. The earlier clamp
  pass covered floating panels but never audited form controls - this
  sweep checked every raw token consumer (40) and capped the six that
  can grow past one line.
- **Playground: dark mode paints the overscroll.** `html` and `body` carry the scheme backgrounds and `theme-color` metas, so iOS rubber-banding no longer flashes white above and below a dark page. (Deployed playground only - no package change.)

### 4.9.0 - 2026-07-30

#### Added

- **Showcase registry: the effects join - the registry now covers the whole library.** Shine border, meteors, marquee, spotlight card, number ticker, text animation and confetti get curated, compile-checked example modules (meteors seeded so the same sky renders every time; text animation fans out to all four functions). With this the 4.9 conversion sweep is complete: every playground component page renders from the shared registry.
- **Showcase registry: the structure set joins.** Card, tabs, table, pagination, breadcrumbs, stepper, button group, user menu and badge get curated, compile-checked example modules, and the form-field registry gains the sliders (range and range-dual are field types) - 24 examples covering composed cards, the three tab styles, honest-sorting tables with people cells, link- and event-mode pagination, steps from plain maps in all three arrangements, fused/split/toolbar/mixed-rail button groups, and the app-shell user menu.
- **`link_type="button"` is now a declared value on `tab` - and safe inside forms.** `Link.a` always rendered it (the playground has used button tabs since the tabs page shipped), but the attr declaration didn't admit it, so every compile emitted value warnings. Declared and documented - and the tab pins `type="button"`, because a bare `