@aceshooting/lyra-ui 7.7.0 → 7.8.1

package/CHANGELOG.md

@@ -1,5 +1,284 @@
 # Changelog
 
+## 7.8.1
+
+### Patch Changes
+
+- d699c7c: Stop letting a throwing `ShadowRoot.activeElement` getter escape a component. `ShadowRoot.activeElement`
+  is not universally safe to read: under happy-dom 20.11.1 — the DOM a large share of consumers get by
+  default from Vitest — that getter _itself_ throws `TypeError: Cannot read properties of undefined
+(reading 'getRootNode')` whenever the document has no active element. Optional chaining was no
+  defence, because `root?.activeElement` only guards `root` being nullish and the throw happens
+  _inside_ the getter, after `?.` has already decided to proceed.
+
+  Because these reads live in `willUpdate()` and in keydown handlers, the symptom was not a failed
+  assertion but an _unhandled rejection_: one downstream suite reported 120 in a single run, all the
+  same stack, from an `<lr-segmented>` re-rendering after its items changed. The suite still passed
+  while the runner exited non-zero, and the stack pointed at library internals rather than anything
+  the consumer wrote.
+
+  Reported against `<lr-segmented>`, but a sweep found the same read at **every** focus-rehoming,
+  roving-tabindex and focus-restoration site in the library — 30 modules across 11 families, including
+  `<lr-tabs>`, `<lr-stepper>`, `<lr-table>`, `<lr-tree>`, `<lr-graph>`, `<lr-combobox>`'s siblings and
+  the shared overlay manager. All of them now read through a new internal helper that returns `null`
+  instead of throwing; `<lr-tree>`'s nested-shadow-root walk was the worst case, reading the raw
+  getter in its _loop condition_ where a guard on the assignment alone would not have helped.
+
+  Returning `null` is the honest answer: a DOM that cannot say what is focused is indistinguishable,
+  for these call sites, from one where nothing is — and every one of them already handles that as the
+  ordinary state, so the guard degrades to skipping focus restoration. Real browsers never take the
+  catch, so behavior there is unchanged.
+
+- 21b1051: Declare `flag-peer.js` in `package.json#sideEffects`, so `<lr-flag>` still resolves images in a
+  production build. `sideEffects` is an explicit allowlist, and every entry in it was derived from a
+  `*.class.ts` file's sibling registration module. `flag-peer.ts` has no `*.class.ts` of its own, so
+  neither the generator nor the completeness check ever visited it and it shipped undeclared. It is a
+  side-effect-only module — a consumer writes a bare `import '.../flag-peer.js'` and reads no export
+  — so any bundler honoring `sideEffects` dropped it outright. `setFlagUrlResolver()` then never ran,
+  `loadFlagUrlResolver()` cached `Promise.resolve(null)`, and every `<lr-flag>` given a
+  `country`/`language` rendered the localized "flag unavailable" alert instead of an image. Silently:
+  that null-resolver path logs nothing, and dev servers don't tree-shake, so it only ever appeared in
+  a built artifact.
+
+  Both scripts now derive `*-peer.ts` and `*-register.ts` modules, plus the per-family
+  `components/<family>/index.ts` barrels, straight from the file tree rather than carrying them over
+  from the previous `package.json` — so a rename or a family move can't strand an entry again. The
+  completeness check fails on the missing `flag-peer` entries before the fix and passes after.
+
+- 899543f: Centre content that a hit-area floor makes narrower than its own box, in `<lr-calendar>`,
+  `<lr-citation-badge>`, `<lr-entity-chip>`, `<lr-rating>` and the `<lr-chart>` / box-plot legend
+  items. These carry the same defect reported against `<lr-widget>`'s view toggle: a flex part with a
+  `min-inline-size: var(--lr-icon-button-size)` floor (a WCAG 2.5.8 tappable-size requirement, not a
+  layout intent) but no `justify-content`, so whenever the content is narrower than that floor the
+  default `justify-content: normal` — resolving to `flex-start` — dumps every pixel of slack on the
+  trailing side.
+
+  `<lr-calendar>` was the most visible: its month-nav buttons hold a single chevron glyph and rendered
+  **8.8px** off centre, sitting right next to a symmetric month title. `<lr-citation-badge>` left its
+  one- or two-digit number hugging the badge's leading edge, and `<lr-entity-chip>` did the same to a
+  short entity label inside an otherwise symmetric pill.
+
+  Adding `justify-content: center` only changes rendering in precisely the buggy case: once content
+  already fills or exceeds the floor there is no slack left to redistribute and the declaration is a
+  no-op, so every component whose content was already wide enough renders exactly as before. The chart
+  legends were checked for the overflow case specifically — long series names wrap rather than
+  overflow, and both legends are wrapping horizontal rows, so per-item centring cannot make a column
+  of items ragged.
+
+  The sweep also cleared roughly ten other parts carrying the same floor where `flex-start` is
+  correct — full-width header and list rows, whose content should start-align regardless.
+
+  Rendered-geometry regression tests (measuring the glyph's centre against its button's) cover
+  `<lr-widget>` and `<lr-calendar>`; both reproduce the offset without the fix, the widget one at the
+  same 4.5px the report cited.
+
+- 8a67993: Stop driving mask alpha from `--lr-color-shadow` in `<lr-segmented>`, `<lr-tabs>`, `<lr-stepper>`,
+  `<lr-timeline>` and `<lr-document-preview>`. All five used `var(--lr-color-shadow)` for the
+  _opaque_ stops of a `mask-image` gradient — 22 references across 12 declarations. A mask reads
+  alpha only, but that token is a documented consumer theming input (`--lr-theme-color-shadow`) whose
+  job is coloring shadows: setting it to something translucent such as `rgb(0 0 0 / 0.25)`, entirely
+  reasonable for a shadow color, silently dropped the mask alpha across the _entire_ element rather
+  than just its edges. Every affected component then rendered uniformly washed out — indistinguishable
+  from a broken disabled state, with nothing pointing back at the shadow token as the cause. It worked
+  only because that token's default happens to be opaque black.
+
+  The opaque stops now use a new `--lr-mask-opaque`, declared in the internal tokens sheet.
+  Deliberately **not** themeable and deliberately not a second alias of the shadow token: "opaque" is
+  not a design decision a consumer tunes — a mask's opaque stop must be opaque by definition — so
+  giving it its own `--lr-theme-*` hook would just reintroduce the same footgun under a new name.
+
+  `<lr-document-preview>`'s determinate progress ring was the least obvious casualty: its mask punches
+  the ring's centre out, so a translucent shadow theme faded the whole ring rather than cutting a hole
+  in it.
+
+  Regression-tested in all five components by rendering under `--lr-theme-color-shadow: rgb(0 0 0 /
+0.25)` and asserting the resolved computed mask, which reproduces `rgba(0, 0, 0, 0.25)` at the
+  opaque stops without the fix.
+
+- 8a67993: Paint the horizontal edge fade only while the track actually overflows, in `<lr-segmented>`,
+  `<lr-tabs>`, `<lr-stepper>`, and `<lr-timeline>`. All four applied their `--lr-scroll-fade-size`
+  `mask-image` unconditionally, described in-code as an intentionally static, observer-free
+  affordance. That is only harmless when there _is_ overflow. On a track that fits, the fade is pure
+  damage: at the `2rem`-per-edge default a two-option `<lr-segmented>` (`Overall | Daily`) is
+  narrower than its own two fades, so both labels rendered half-transparent and the control read as
+  permanently disabled; a short `<lr-tabs>` row dimmed its first and last tab for no reason.
+
+  A new internal `ScrollOverflowController` measures `scrollWidth` vs `clientWidth` and toggles a
+  `data-scroll-overflow` attribute on the track (inside the shadow root — not consumer-visible DOM),
+  which now gates each mask rule. It re-measures from two sources, because they catch different
+  changes: a `ResizeObserver` on the track for container resizes, and the host's own update cycle for
+  content changes, which need not alter the track's border box at all. Overflowing tracks keep
+  exactly their previous rendering.
+
+  Note for anyone spying on `ResizeObserver` construction: `<lr-stepper>` now arms one of its own
+  regardless of the `orientationBreakpoint` feature.
+
+- 899543f: Center the glyph in an icon-only `<lr-widget>` view toggle. `[part="view-toggle"]` set
+  `align-items: center` but no `justify-content`, unlike the sibling `collapse-button` /
+  `fullscreen-button` rules, which set both. `min-inline-size` floors the pill at the square
+  icon-button size, so a 13px glyph inside a 40px pill has slack that the default
+  `justify-content: normal` (→ `flex-start`) dumps entirely on the trailing side — measured 4.5px off
+  true center once the asymmetric inline padding is counted, and plainly visible as an off-center
+  icon in a round toggle. A labeled toggle was never affected: its content already fills a pill that
+  sizes to fit.
+
+## 7.8.0
+
+### Minor Changes
+
+- 2c3934b: `<lr-embedding-explorer>`'s `height` property now actually sizes the plot. It was rendered as an
+  SVG `height` presentation attribute while the component's own stylesheet declared
+  `[part='plot'] { block-size: auto }` — and any stylesheet declaration outranks a presentation
+  attribute, so the property was inert at every value, including its documented `360px` default: the
+  plot always sized itself from the `viewBox` aspect ratio instead.
+
+  `height` is now published on the host as the new `--lr-embedding-explorer-height` custom property,
+  which `[part='plot']`'s `block-size` reads. Consequences worth knowing before upgrading:
+
+  - The default `height="360px"` now takes effect, so a plot wider than 640px is no longer as tall as
+    its allocation implies. Set `height="auto"` to keep the previous aspect-ratio-preserved sizing.
+  - A value the browser cannot parse as a `block-size` is dropped rather than applied, leaving the
+    `auto` behavior instead of collapsing the plot.
+  - A consumer's own `::part(plot) { block-size: ... }` rule still overrides `height`, and the
+    narrow-allocation `min-block-size` floor still raises it.
+
+- 2c3934b: `<lr-memory-panel>` no longer strands keyboard focus when a row action opens its confirmation step.
+  Activating "Add to long-term memory", "Remove", or "Forget all" destroys the button that had focus,
+  and nothing moved focus into the `lr-confirm-bar` that replaces it, so focus fell back to `<body>`:
+  a keyboard user was dumped at the top of the page with nothing announced, and had to re-tab through
+  the whole document to reach the confirmation they had just opened. Focus now moves into the
+  confirmation (its Deny control -- the safe action -- falling back to the bar's status element), and
+  is handed back to the row (or to the "Forget all" control) once the decision resolves. Pressing
+  Escape while the confirmation holds focus now cancels it exactly like pressing Deny: no event is
+  emitted, focus returns the same way, and the key does not propagate past the panel.
+- 2c3934b: `<lr-notebook-viewer>`: `searchNext()` and `searchPrevious()` now resolve `true` when the active
+  match moved and `false` when there was nothing to move to, matching the shared viewer search
+  contract (`LyraTextViewerTarget`) that every other searchable viewer already honors. They
+  previously returned nothing, so a find-in-page host driving several viewers polymorphically —
+  `if (await viewer.searchNext()) { ... }`, or awaiting the call before reading its own match
+  counter — got `undefined` from the notebook viewer alone and took its falsy "no more matches"
+  branch on every press, even mid-notebook.
+
+  This is an additive widening: the methods return a resolved promise instead of nothing, and callers
+  that ignored the return value are unaffected. `search()` already resolved the match count and is
+  unchanged.
+
+- 80e0ef1: Repair four regressions left by earlier fixes:
+
+  - `<lr-radio-group>`: arrow-key selection now emits `input` and `change` alongside `lr-change`, as
+    click and Space already did and as native `<input type=radio>` does. The earlier fix for a
+    duplicate `lr-change` had left the keyboard path emitting only the group event, so a consumer
+    bound to the native-mirroring events silently missed every keyboard selection.
+  - `<lr-progress-ring>`: an unslotted ring is named from the localized fallback again. Its slot's
+    fallback content is the formatted percent, and `assignedNodes({flatten:true})` returns fallback
+    children when nothing is assigned, so the control had been naming itself "40%" and no
+    `registerLyraLocale()` override could reach it.
+  - `<lr-tour>`: opening a detached tour no longer locks scroll on the document or installs a global
+    Escape handler with nothing visible, matching the guard `<lr-dialog>` already had.
+  - `text-quote` anchors and highlights now case-fold with the component's locale in
+    `<lr-docx-viewer>`, `<lr-pdf-viewer>`, `<lr-markdown>` and `<lr-markdown-core>`. Under `lang="tr"`
+    a quote of "istanbul" silently failed to match "İSTANBUL" in these four while resolving correctly
+    in every viewer built on the shared text-viewer mixin.
+
+- d8a026d: Give `<lr-table>`'s sorted column header an opaque default fill. The header is `position: sticky`
+  and the sorted-state rule defaulted to `transparent`, so in any height-capped table the body rows
+  scrolled visibly through the sorted column's header cell.
+
+  Give `<lr-pdf-viewer>`'s toolbar buttons a hover fill that differs from the toolbar behind them —
+  the rule existed but resolved to the toolbar's own opaque token, so hovering produced no visual
+  change at all. Retunable via the new `--lr-pdf-viewer-toolbar-button-hover-bg`.
+
+  Correct three `<lr-chat-message>` snippets in the authored reference that used `role="user"`. The
+  property reflects to `data-role`, so `role` was never observed: consumers copying those examples
+  got a message rendered as the default `assistant`, plus an invalid ARIA role token in the DOM.
+
+- 2c3934b: `<lr-xml-viewer>`, `<lr-av-player>`, and `<lr-terminal>` now resolve a boolean from `searchNext()`
+  and `searchPrevious()`, matching the shared `LyraTextViewerTarget` search contract that
+  `search()` already followed on all three. They returned `void`, so a host driving several
+  searchable components through that one typed surface — `if (await viewer.searchNext())` — read
+  `undefined` and took the "nothing to move to" branch on every press.
+
+### Patch Changes
+
+- ed7f463: Stop retaining one live `Range` per search match in every text viewer. `<lr-archive-viewer>`,
+  `<lr-calendar-viewer>`, `<lr-contact-viewer>`, `<lr-email-viewer>`, `<lr-geojson-view>`,
+  `<lr-html-viewer>`, `<lr-include>`, and `<lr-pptx-viewer>` share a search mixin that held a live
+  `Range` for every match. The engine revalidates each retained `Range` on every DOM mutation in its
+  document, so a short query over a long document made every later mutation dramatically slower.
+  Matches are now kept as inert offsets, and only a bounded window around the active match is
+  materialized and painted. `matchCount`, `searchNext()`, and `searchPrevious()` still cover every
+  match.
+- 2d149dc: `<lr-dashboard-grid>` no longer starts a cell drag when the pointer lands on a button, link, or
+  input inside the cell. The guard compared a slotted light-DOM control against a shadow-root wrapper
+  with `contains()`, which never crosses the slot boundary, so it could never fire and every control
+  click inside a draggable cell dragged the cell instead of activating the control.
+
+  `<lr-tree>` no longer throws on the first arrow key when a `<lr-tree-node>` is written declaratively
+  into its documented slot. `item` is `attribute: false`, so such a node has none until a host assigns
+  one, and the keyboard handler read `item.id` unguarded.
+
+- 9083c9b: `<lr-document-library>` now sorts its Updated column chronologically. It ordered rows correctly by
+  timestamp itself, then handed the composed `<lr-table>` both those rows and a `sortKey` without
+  `sort-mode="server"`, so the table sorted them a second time in client mode — from the column's
+  rendered output, which is a _formatted_ date. The result was alphabetical by month name.
+- 2c3934b: `<lr-env-list>` no longer paints its screen-reader-only "Value hidden" announcement as visible text
+  beside the mask. The template emitted `class="sr-only"` but the component never adopted the shared
+  stylesheet that defines that class, and no rule in `LyraElement.styles` supplies it.
+- 2c3934b: Hide the anchor-announcement live region in `<lr-include>` and `<lr-pptx-viewer>`.
+
+  Both viewers render the shared anchor-target mixin's `role="status"` live region, which the mixin
+  marks up with `class="sr-only"`, but neither component's shadow stylesheet defined that class. The
+  region therefore laid out as an ordinary block, so the first anchor jump (or a failed one) painted
+  its localized announcement — "Jumped to highlighted passage." / "Passage not found in this
+  document." — as visible body text: beside the transcluded fragment for `<lr-include>`, and as an
+  extra row under the fidelity notice for `<lr-pptx-viewer>`. The announcement is now visually hidden
+  and screen-reader-only, matching every other viewer that adopts the same mixin.
+
+- 0e2dbf3: `LyraElement`'s internal `scheduleAfterUpdate()` now coalesces per key instead of collapsing every
+  caller in an update cycle onto one slot. It tracked pending work in a single boolean, so the second
+  caller in a cycle early-returned and its callback was dropped and never replayed — a component that
+  scheduled two genuinely different pieces of after-update work silently lost one of them. Repeated
+  schedules under the same key still collapse to one run, so the load path keeps producing one fetch
+  per cycle rather than one per property write.
+- 2cd5fb5: `<lr-artifact-panel>`, `<lr-commit-card>`, `<lr-heatmap>`, `<lr-query-builder>`, `<lr-tree>`, and
+  `<lr-word-cloud>` now format the numbers they interpolate into localized strings with the effective
+  locale. `localize()` substitutes values with a bare `String(value)` and does no number formatting,
+  so these rendered Western digits inside otherwise fully-translated sentences — under a locale using
+  its own numbering system (`ar-u-nu-arab`, `hi-u-nu-deva`, …) a single announcement mixed two digit
+  sets.
+
+  `<lr-attachment-chip>` also no longer falls back to an empty `src` on its thumbnail `<img>`; an
+  empty `src` is a valid URL that resolves against the document, so it would make the browser
+  re-request the page as an image.
+
+- df4dac8: Reject non-integer index segments in anchor resolution. A range-only guard (`i < 0 || i >= len`)
+  does not reject `NaN` (both comparisons are false) or a fractional index, so `<lr-xml-viewer>`
+  reported `lr-anchor-result { found: true }` and announced "Jumped to…" for a `node-path` that
+  matched nothing, and a non-trailing bad segment threw — rejecting `scrollToAnchor()` so
+  `lr-anchor-result` never fired at all, and surfacing as an unhandled rejection on the declarative
+  `anchor` path. `<lr-notebook-viewer>` had the same false-positive shape, and
+  `<lr-virtual-list>.scrollToIndex(NaN)` silently scrolled the list to the top.
+
+  `DocumentAnchorTarget` now also degrades a throwing `applyAnchor()` to "not resolved" instead of
+  letting it reject, so the mixin keeps its documented promise of always reporting a definite result.
+
+- c76ebf8: Fix `<lr-mcp-app>`'s remote `src` mode, which never loaded: binding `srcdoc` to an empty string
+  still produced a _present_ `srcdoc=""` attribute, and the HTML spec's iframe processing branches on
+  that attribute's presence, so the frame navigated to `about:srcdoc` and ignored `src` entirely
+  (while still firing `lr-mcp-ready`). The same empty-string-vs-absent shape is fixed in
+  `<lr-av-player>` (a bare player painted a "Failed to load the media" alert before a `src` was set)
+  and `<lr-zoomable-frame>` (a rejected `src` rendered a broken-image glyph).
+
+  Validate consumer-supplied CSS lengths before they reach an inline style declaration list, so a
+  crafted value can no longer inject extra declarations: `<lr-stack-trace>`'s `max-height`,
+  `<lr-code-block>`/`<lr-code-block-core>`'s `max-height`, `<lr-table>`'s column
+  `width`/`minWidth`/`maxWidth`, and `<lr-browser-frame>`'s agent-supplied ping coordinates (which are
+  now also clamped to the documented 0-100 range instead of serializing `NaN`).
+
+- ed7f463: `<lr-virtual-list>` no longer rescans the whole `items` array on every scroll frame to resolve
+  `active-id`. The lookup is now memoized on the `items`/`active-id`/`keyFunction` identities, so
+  scrolling a large list stops calling `keyFunction` once per item per frame.
+
 ## 7.7.0
 
 ### Minor Changes