@ids-group-ltd/ids-design-system 0.9.1 → 0.10.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -8,6 +8,955 @@ While the package is on `0.x`, minor releases may contain breaking changes.
8
8
 
9
9
  ## Unreleased
10
10
 
11
+ ## 0.10.1 — 2026-08-20
12
+
13
+ ### Removed
14
+
15
+ - **`--opacity-muted` is gone from the token scale.** It was the last token in the file that no
16
+ component read: the disabled states moved onto `--text-disabled` + `--opacity-disabled` in
17
+ 0.10.0, and it had been sitting there since, labelled a legacy alias. Keeping a dead number in a
18
+ purpose-named scale is worse than removing it — it invites exactly the override that 0.10.0's
19
+ DELETE-LIST had to warn about twice, where a consumer retunes it on `ds-button` /
20
+ `ds-icon-button` and gets a silent no-op. **Nothing in the library changes.** If you set it, drop
21
+ the declaration; if you READ it in your own CSS, inline the 0.6 or move to the token that names
22
+ the job — `--opacity-disabled` (0.5) for a disabled state where opacity is the only signal,
23
+ `--opacity-text-disabled` (0.38) for disabled text. The comment above the block claims every
24
+ token in it has a real consumer; that is true again.
25
+
26
+ ### Added
27
+
28
+ - **`ds-segmented` joins the form systems, and gains a label and a message row.** It carried a value
29
+ and no form could bind it — no `NG_VALUE_ACCESSOR`, so `[(ngModel)]`, a `formControlName` and
30
+ signal forms all slid off it. It now extends `DsControlValueAccessor` and resolves its message
31
+ through the same `resolveFieldMessage` as every other field, mirroring `ds-radio-group`: `label`,
32
+ `hint`, `error`, `warning` and `errors` all behave as they do elsewhere, and `disabled` folds
33
+ together with the form's own disabled state. **The DOM grew a level to fit them:** the pill moved
34
+ off `:host` onto an inner `.track`, so the host can stack label, pill and message. The pill's own
35
+ tokens stayed on `:host` and inherit down, so a token override is unaffected — but a
36
+ `::ng-deep` rule written against the host AS the pill (a background, a border-radius, a grid) now
37
+ targets the stack instead, and wants `.track`.
38
+
39
+ - **`invalid` paints the error state only — it no longer promotes `hint` to the error message.**
40
+ Every state carries its own text now, and `hint` is never recoloured into somebody else's state.
41
+ **BREAKING:** `[invalid]="true" hint="Must be before end date"` used to render that hint in red,
42
+ through `role="alert"`. It now renders a red border with a **grey** hint. Use
43
+ `error="Must be before end date"` instead — the state and the message come together.
44
+ **Why, beyond tidiness.** The idiom lost the actual hint: when both were set, the help text the
45
+ author wrote was replaced by the error rather than shown alongside it. It also announced help text
46
+ as an alert, and it made a template unreadable — whether `hint="…"` was a hint or an error depended
47
+ on a different attribute. It was the last input that borrowed another one's text; `success` stopped
48
+ doing it in the same release.
49
+ **`invalid` is not going away and still has its own job:** the error state with NO message, for when
50
+ the text lives elsewhere — a form-level error summary, or one message covering a whole
51
+ `ds-field-group` — plus `aria-invalid` for that case. Carbon and Material split the boolean from
52
+ the text the same way.
53
+ Our own docs taught the old idiom on 13 components (14 of 19 `[invalid]` bindings also bound
54
+ `hint`); all of them now use `error`.
55
+
56
+ - **A form model can now raise a WARNING, not only an error** — `DS_FIELD_WARNING_KIND`, newly
57
+ exported. A validation error whose `kind` is `'warning'` paints the field's warning state and shows
58
+ its message, but leaves the control **valid**: `aria-invalid` stays off and a submit guard reading
59
+ validity is not blocked.
60
+ **Why it had to ride inside `errors`.** Angular's `AbstractControl` has no channel for an advisory
61
+ remark — it exposes `errors` and the validity derived from them, and nothing else. So there was no
62
+ way for a form to say "this is worth mentioning but not wrong": `warning` and `success` could only
63
+ be bound by hand from the consumer's own state.
64
+ **It works from both form systems with no new plumbing**, which is the point. A reactive or
65
+ template-driven validator returning `{ warning: 'Long runs of capitals are hard to scan' }` has its
66
+ KEY read as the kind; signal forms assigns its own `ValidationError[]`, which already carries a
67
+ `kind`. Both land in the same `errors` signal.
68
+ Ranked below the explicit `warning` input — the consumer's own string wins, exactly as `error` beats
69
+ `errors` — and shown only after the user has interacted, like any other validation message.
70
+ **`DsFieldError` and `DsFieldErrorsInput` are exported too**, so a typed validator can be written
71
+ against the shape the input accepts. Reserved-kind caveat: a validator already using the key
72
+ `warning` for a BLOCKING error now reads as advisory. Rename that key if you have one.
73
+
74
+ - **`success` exists on all nine bordered fields, and it is now a STRING that carries its own text.**
75
+ Added to `ds-textarea`, `ds-select`, `ds-multi-select`, `ds-autocomplete`, `ds-date-picker`,
76
+ `ds-time-picker`, `ds-otp-input` and `ds-color-picker` — it was on `ds-input` alone, so a form could
77
+ not use the state consistently. The three box-less controls (`ds-checkbox`, `ds-radio-group`,
78
+ `ds-switch`) and `ds-slider` deliberately do NOT get it: they have no box for a border to appear on,
79
+ so the state would show in the message line alone and read as a different treatment.
80
+ **BREAKING for `ds-input`, the one field that already had it:** `success` was
81
+ `input(false, { transform: booleanAttribute })` and is now `input<string>('')`. Replace
82
+ `[success]="true" hint="Available"` with `success="Available"`.
83
+ **Why the type changed rather than staying a flag.** It was the only state that could not say
84
+ anything: `error` and `warning` carry their own text, and `success` was a boolean that recoloured
85
+ the `hint` green. So `success` was the one state whose message had to be smuggled through a
86
+ different input, and a reader of the template could not tell a green hint from a grey one. It is now
87
+ a fourth rung in the same ladder — `error` → `errors` → `warning` → **`success`** → `hint` — and
88
+ reads exactly like `warning`: the text IS the state, no second binding.
89
+ **What this removes:** a success border with no message, which is no longer expressible — exactly as
90
+ it never was for `warning`. `invalid` still exists as the flag partner of `error`, for the
91
+ `[invalid]="…" [hint]="…"` idiom that predates `error`.
92
+
93
+ - **`ds-list-item` can hold its own controls: `[leading]` and `[trailing]` render OUTSIDE the row's
94
+ button.** A row with a checkbox AND a row-level action had no shape here — every projected node went
95
+ inside `<button class="activator">`, where a nested control is invalid HTML, its click bubbles into
96
+ `activated`, and a screen reader folds it into the button's name and never announces its checked
97
+ state. Dropping `clickable` was no answer either: `cursor`, `:hover`, `:focus-visible` and
98
+ `aria-current` all hung off that button. The row is now the flex container and the button wraps only
99
+ the text, so the master–detail shape works — a checkbox per row for a bulk action, the row itself
100
+ choosing what a detail pane shows, two independent targets.
101
+ **This stays a `role="list"`, deliberately not a `listbox`:** the ARIA spec is explicit that a
102
+ `listbox` option cannot hold interactive content, so a real listbox would have made the pattern
103
+ impossible rather than easier. `role="listitem"` carries no such restriction.
104
+ **`selected` also works properly now.** It set `aria-current` on the inner button, so it did nothing
105
+ at all unless `clickable` was on; `aria-current` sits on the host, which is the row that is current.
106
+ **BREAKING for CSS overrides only** — no API changed. `.activator` no longer carries the row's
107
+ padding, gap or hover tint: those moved to the host. A `::ng-deep .activator { padding: … }` now pads
108
+ the text alone, and a `::ng-deep .activator:hover` no longer paints the full row.
109
+
110
+ - **Per-side padding on `ds-modal` / `ds-modal-card`.** Each region resolved ONE `pad-y` applied to
111
+ both of its edges, so a layout whose header pads only at the top and whose actions row pads only
112
+ at the bottom — the body's own padding carrying the whole gap between them — could not be said
113
+ through the API at all. mm-broker has held three `!important` rules on `.modal > .header` /
114
+ `> .body` / `> .actions` for exactly this, reaching past the component into its internal class
115
+ names, and `DS_CONFIG.modal.dividers` could not retire them. Six new tokens —
116
+ `--ds-modal-header-pad-top` / `-pad-bottom` and the same for `body` and `actions` — each fall back
117
+ to their region's existing `pad-y`, so **nothing moves unless you set one**, and `--ds-modal-*-pad-y`
118
+ keeps working as the shorthand for both edges.
119
+
120
+ - **`ds-table`'s own icon buttons are themeable.** The row-action kebab and the column-picker trigger
121
+ hardcoded `--icon-default`, `--icon-strong` on hover and an `--accent` focus outline, and neither is a
122
+ `ds-icon-button`, so the icon-button tokens never reached them — a consumer whose design calls for a
123
+ teal row menu had no supported way to get one short of repainting `--icon-default` for the whole app.
124
+ Three resolvers now cover both: `--ds-table-icon-btn-color`, `--ds-table-icon-btn-color-hover` and
125
+ `--ds-table-icon-btn-focus-color`, defaulting to exactly what was hardcoded, so nothing moves unless
126
+ you set them.
127
+
128
+ ### Fixed
129
+
130
+ - **`ds-field-message`'s docs claimed `state` picks an icon. It never did.** The input's docblock
131
+ said it "decides colour and icon" and `icon`'s said it "overrides the glyph the `state` would
132
+ pick" — but `icon` is a plain `input<IconName | null>(null)` with no per-state default, so a
133
+ consumer following the reference got a colour and no glyph. **Nothing in the library changes**: the
134
+ JSDoc and the generated API reference now say what the code does — `state` sets the colour, and a
135
+ field that wants an icon passes one.
136
+
137
+ - **A nullable binding on a message input no longer throws.** `hint`, `error`, `warning` and
138
+ `success` took a bare `string`, and the resolver called `.trim()` on whatever arrived — so
139
+ `[error]="form.errors?.message"` put `null` in and crashed the render with a `TypeError` the
140
+ moment the expression was empty. All four now transform at the input boundary, absorbing
141
+ `null`/`undefined` to `''` and trimming, so a blank binding simply means "no message". **You can
142
+ now delete** the `?? ''` you wrapped such a binding in.
143
+
144
+ - **A `ds-menu` inside a `ds-popover` clips its rows to the panel's rounded corners.** The first and
145
+ last row's fill — a hover tint, or the red of a destructive item — painted through the panel's
146
+ corner, so a destructive last item showed **two square red corners under a rounded popover**.
147
+ **The cause was not a missing clip; it was a clip of the wrong shape.** `ds-menu` has
148
+ `overflow: hidden`, and that is what rounds the edge rows for a bare menu. But its
149
+ `:host-context(.popover)` block flattened `border-radius` to `--radius-flat` while leaving `overflow`
150
+ in force, so inside a popover the menu clipped its rows **square** and the panel's 12px corner
151
+ showed that square fill through it. The popover now publishes `--ds-menu-clip-radius` — its own
152
+ radius less the border it contains — and the menu clips to that, concentric at 11px inside a 12px
153
+ panel. A brand overriding `--ds-popover-radius` is followed automatically.
154
+ **Not fixed with `overflow: hidden` on the panel**, which was the obvious move: a submenu is a CDK
155
+ overlay portaled to `body`, so clipping the panel would not have reached it, and the panel's clip
156
+ would have cut an edge row's focus ring — the same defect fixed in `ds-filter-bar` this month.
157
+ Nothing else changes: the item radius token `--ds-menu-item-radius` still defaults to 0, because the
158
+ MENU owns the edge corners, not the row.
159
+
160
+ - **Clearing or replacing `ds-date-picker`'s `value` from outside now clears the picker's own typing
161
+ error.** `typedInvalid` / `rangeInvalid` were set by the typing paths and cleared only by those and
162
+ `applyPreset`, so `value.set(null)` re-rendered the field to empty and left **"Enter a valid date
163
+ (MM/DD/YYYY)" standing over it**. `[error]` was no help: a consumer who has just cleared the field
164
+ has no error of its own to bind, and an empty string falls straight back to the picker's.
165
+ The `value → text` effect now clears both flags. It is the right place because every typing path
166
+ holds an `editing` guard while it runs, so reaching the effect means the value did NOT come from the
167
+ keyboard — and a leftover error no longer describes what the field is about to show. Verified that
168
+ neither typing path writes `value` on the invalid branch, so blur validation still stands.
169
+ **This retires two consumer workarounds.** One app was tearing down a whole modal body on `closed`
170
+ purely to get a fresh picker — and that had a hole, since reopening inside the 300 ms leave animation
171
+ cancels the close, `closed` never fires and the stale error survived. A second dialog had no
172
+ workaround at all, so a mistyped date of birth read "Enter a valid date" over an empty field the next
173
+ time it opened.
174
+
175
+ - **A disabled `ds-button` with `variant="ghost"`, `variant="text"` or `variant="outline"` mixes its
176
+ label toward the surface instead of taking the shared fixed-lightness recipe.** New default:
177
+ `color-mix(in srgb, <label> 40%, var(--surface-default))`. These three share one accent source and
178
+ have no fill that could carry the state, so the label is the whole signal — and the shared recipe
179
+ was failing it, badly in dark. Measured against the enabled label: light **2.54 → 3.41:1**, dark
180
+ **1.11 → 2.91:1**. `outline` does have a border, but its disabled border resolves to only
181
+ **1.30:1** against the page — less than its own *enabled* border at 1.49:1 — so it never covered
182
+ for the label.
183
+ **The dark number is the real defect.** `--primary-strong` there is already light (OKLCH L 67.9%),
184
+ so a recipe pinning the disabled label at L 70% landed two points from the enabled one and the two
185
+ differed in chroma alone. A fixed lightness cannot serve both themes: in light, disabled is
186
+ lighter than the accent; in dark the accent is already light, so disabled has to go darker. The
187
+ surface knows that direction and a constant does not — which is the whole reason the mix target is
188
+ `--surface-default` rather than a number.
189
+ **The trade:** the disabled label is fainter against its own ground — **2.00:1** light, 1.99:1
190
+ dark, where the old recipe gave 2.69 / 6.45. Disabled text is exempt from WCAG 1.4.3 and 1.4.11
191
+ and Material ships its equivalent at 1.93:1, so this is inside practice. Chroma rises with the
192
+ mix (0.036 → 0.099 light), so the disabled label stays tinted rather than going grey; that was
193
+ chosen deliberately over a desaturating variant after looking at both.
194
+ **`solid` and `tonal` are unchanged**, deliberately: a white label mixed into a white surface does
195
+ not move, and tonal's disabled fill already carries the state on its own. **Known limitation:** the
196
+ mix target is `--surface-default`, a stand-in for "whatever is behind" — CSS cannot read the actual
197
+ backdrop, so on a tinted ground (a card at `--surface-secondary`, an alert) the disabled label is
198
+ computed against the page colour rather than that ground. Consumers who set
199
+ `--ds-button-label-color-disabled` are unaffected.
200
+ ### Changed
201
+
202
+ - **A disabled `ds-icon-button variant="standard"` mixes its glyph toward the surface too.** It is
203
+ the only icon-button variant with neither a fill nor a border, so the glyph's colour is the whole
204
+ disabled signal — and the shared fixed-lightness recipe could not carry it in dark, where the
205
+ enabled icon is already light (OKLCH L 80%) and a disabled one pinned at L 70% sat **1.46:1** away.
206
+ The new default is `color-mix(in srgb, <icon> 40%, var(--surface-default))`, the same mix the
207
+ borderless button variants take, which picks its direction per theme: light **4.17 → 5.19:1**, dark
208
+ **1.46 → 3.58:1**. Scoped to `standard` deliberately — `filled`'s icon is `--c-on`, and white
209
+ mixed into a white surface does not move at all. `--ds-icon-button-icon-color-disabled` is the same
210
+ hook as before and still wins; only the default under it changed.
211
+
212
+ - **If you EXTEND a DS field component, read this before trusting `class-surface`.** The report
213
+ against 0.10.0 says fourteen field components LOST eight members each — `hint`, `error`,
214
+ `warning`, `errors`, `invalid`, `touched`, `reserveMessage` and `reserveMessageResolved`.
215
+ **Nothing was removed.** Those inputs were declared once per component and are now declared once in
216
+ total, on a new internal `DsFieldBase`, which every field reaches either directly or through
217
+ `DsControlValueAccessor` / `DsBooleanControlValueAccessor`. All fourteen still carry all eight by
218
+ inheritance, so a template binding, a TS read and an existing `override` all behave exactly as
219
+ before. The tool counts declarations per FILE, and `migrations/AUTHORING.md` already names a member
220
+ that moved to a base class as one of the two things it cannot see — this release is that case.
221
+ What genuinely grew is smaller: **eight bordered fields gained `success`** (`ds-autocomplete`,
222
+ `ds-color-picker`, `ds-date-picker`, `ds-multi-select`, `ds-otp-input`, `ds-select`,
223
+ `ds-textarea`, `ds-time-picker`), `ds-input` having had it all along. `ds-segmented`
224
+ is the tenth grown class and the one with real new surface — joining the form systems gave it
225
+ `label`, `disabled`, `isDisabled`, `message`, `invalidResolved`, `fieldMessage`, `hintId` and
226
+ `uid`. A subclass declaring
227
+ `success` on one of those eight now fails with TS4114; declaring any of the eight moved names still
228
+ collides exactly as it did, because an inherited member collides the same way an own one does.
229
+ `DsFieldBase` is internal and deliberately not exported — it holds only the inputs every field has,
230
+ which is why `success`, `label`, `required` and `optional` stayed on the components.
231
+
232
+ ## 0.10.0 — 2026-08-18
233
+
234
+ ### Added
235
+
236
+ - **A capped chip trigger no longer grows a second line on a narrow screen.** Found by Stas testing the
237
+ docs on a phone: `[maxVisibleChips]="2"` honoured the count, ran out of width, and wrapped the text
238
+ input onto its own line — the exact growth the cap exists to prevent. **A number is now a CEILING**
239
+ ("at most N"), measured like `"fit"`, so a 412px trigger shows one chip and `+3 more` on one line
240
+ instead of two chips over two lines. `null` is unchanged and still the only mode that skips measuring.
241
+ Three faults sat behind it, all fixed: the trigger kept `flex-wrap: wrap` while capped, so the row
242
+ could never shrink — it just moved the input down; a squeezed chip wrapped its own label and stood
243
+ 46px tall beside 29px neighbours; and the measurement's width cache only filled on a render holding
244
+ EVERY chip, which a numeric cap never produces — so the fit silently fell back to "show them all".
245
+ The `+N more` marker also keeps its room whenever the cap hides anything, instead of being clipped to
246
+ `+2 mor`.
247
+
248
+ - **A `TableColumn` can say `width: 'content'`.** The column sizes to its widest cell and stops
249
+ wrapping, so the slack goes to the columns that should absorb it. An auto-layout table spreads
250
+ leftover width evenly, which is why a status or a figure ends up as wide as its own header while the
251
+ name column that should take the space does not — and a pixel `width` cannot express that, since
252
+ pinning one column leaves the rest spreading. Measured in the docs demo: two `'content'` columns come
253
+ out at 89px and 94px while the two ordinary ones take 438px and 530px of the same 1152px table.
254
+ **You can now delete** the `::ng-deep ds-table { .table th, .table td { width: 1%; white-space: nowrap } }`
255
+ block — the last reach into the table's markup a consumer had left, still marked in their tree as
256
+ `TODO(ds-shrink-to-content)`. Dragging a column's edge still wins, because that is a deliberate
257
+ pixel width — but note it wins for the WHOLE table at once: on a `resizableColumns` table the first
258
+ drag OR resize keystroke anywhere locks a measured pixel width onto every column that has not got
259
+ one, `'content'` columns included. `Home` on a column's own handle hands that one back; the column
260
+ panel's "Reset to default" button and clearing the public two-way `columnWidths` hand back all of
261
+ them. A `'content'`
262
+ column also has no clip width, so `truncate` on it is a no-op rather than an ellipsis against 240px.
263
+ The `title` attribute and the click-to-popover handler are still emitted on such a cell, but the
264
+ popover cannot open: it opens only when the cell measures as clipped, and this one never does.
265
+
266
+ - **`ds-table` takes `[loading]="true"` and picks the treatment itself** — skeleton rows while it has
267
+ nothing on screen, overlay once there are rows. `'skeleton'` and `'overlay'` still pin either one
268
+ deliberately, and `false` is unchanged. Which of the two is right is a function of the row count,
269
+ which the table already holds and every consumer was restating: **you can now delete** the
270
+ `items().length > 0 ? 'overlay' : 'skeleton'` computed beside each list — one consumer feature had
271
+ three copies. The rule is not a taste call, which is why it belongs here: replacing visible rows with
272
+ placeholders throws away what the reader is looking at, and dimming an empty card announces nothing
273
+ about what is coming, so a hardcoded `'overlay'` makes a first load read as an empty table that
274
+ flickers.
275
+
276
+ - **`ds-table` has a toolbar row you can put things in — one `[toolbar]` slot.** The component had two
277
+ projection slots in total (`[empty]`, `[bulk-actions]`) and its toolbar existed only while
278
+ `exportButton` was on, so a control that acts on THIS table — a search box, a status filter, a refresh
279
+ — had nowhere to sit next to Export. A consumer forked the table for exactly this; retiring the fork
280
+ pushed two screens' filter rows out of the table and onto the page, where a search + status pair reads
281
+ as page furniture one card away from the rows it narrows. **Project a handful of controls and the row
282
+ lays them out** (flex, wrap, 8px gap), so the ordinary case needs no CSS of yours; **project one
283
+ element of your own** and it stretches to fill the slot, so your layout is the only one in play — the
284
+ slot is the row minus the Export button, and the whole row when `exportButton` is off. The row
285
+ **collapses to nothing** when nothing filled it and `exportButton` is off, so a table that uses
286
+ neither is pixel-identical — the `div` is still in the DOM, hidden in CSS, so a
287
+ `ds-table > :first-child` selector or a DOM snapshot still sees it. Filters that belong to the PAGE
288
+ still belong in a `ds-filter-bar` above it.
289
+
290
+ - **`runExport()`, `exporting()` and `exportStatus()` are public**, so the Export button can live in a
291
+ page header, a filter bar or the new toolbar slot and still run the library's path — the
292
+ `exportFetcher` call, the download, the in-flight state and the live-region announcement. Until now the
293
+ only public export method was `exportCsv()`, which serialises the rows the table HOLDS: on a
294
+ server-paged grid that is one page, so a consumer's own button silently exported a fraction of the
295
+ data. Leave `exportButton` off and drive it yourself:
296
+ `<ds-button [loading]="table.exporting()" (pressed)="table.runExport()">`. That pairing wants an
297
+ `exportFetcher`: without one, `runExport()` falls through to `exportCsv()` synchronously and never
298
+ touches `exporting`/`exportStatus`, so an external button bound to them shows no progress at all.
299
+ The two signals are read-only — a subclass that used to `.set()` them (they were protected writables)
300
+ drives them through `runExport()` now.
301
+
302
+ - **`provideDsConfig` answers seven more inputs in this entry alone, so an app stops restating
303
+ itself.** Five of the release's eight new sections — `avatar`, `brand` and `modal` have their own
304
+ entries:
305
+ `control.size` (honoured by button, icon-button, input, textarea, select, autocomplete, search,
306
+ date/time picker, number stepper, checkbox, radio, switch and field-group), `card.elevation` and
307
+ `card.padding`, `table.stickyScrollbar` and `table.bordered`, `icon.size`, and `badge.withDot`.
308
+ Precedence is **input → config → built-in**, so a screen that differs still binds the attribute and
309
+ wins. **No default moved** — an app that provides nothing renders exactly as before, with one edge
310
+ worth knowing: the booleans that swapped `booleanAttribute` for a nullable transform
311
+ (`table.bordered`, `table.stickyScrollbar`, `badge.withDot`, and `dividers` on both modal components)
312
+ used to read a binding evaluating to `null`/`undefined` as `false`, and now read it as *unset* — so it
313
+ falls through to the config and then the built-in, which for `dividers` is `true`. A literal
314
+ `[bordered]="false"` is unaffected. **You can now
315
+ delete** the attribute an app repeats on every element: one consumer binds `[stickyScrollbar]="true"`
316
+ on all 35 of its tables and `size="sm"` on 146 buttons, which is a house default written 146 times,
317
+ and each new element is one forgotten binding away from looking wrong. `control.size` is a different
318
+ axis from `[data-density]`: the tier rewrites hit targets and padding in CSS for a route, this picks
319
+ which of the three declared sizes a control renders at, and an app can use both. Documented as one
320
+ page under Foundations → Theming rather than in scraps across eight component pages.
321
+
322
+ - **Option icons on every picker, and one shape behind all three.** `ds-multi-select` and
323
+ `ds-autocomplete` now render an option's `icon` before its label at 16px, which only `ds-select`
324
+ did. The three option types share a new exported `DsOption<T>` — they were three near-identical
325
+ interfaces, which is exactly why `icon` existed on one picker for a release. Each keeps its own
326
+ name (`SelectOption`, `MultiSelectOption`, `AutocompleteOption`), so nothing you import changes.
327
+ From the same shared shape `AutocompleteOption` gains `disabled?` and `icon?`, `MultiSelectOption`
328
+ gains `icon?`, and `SelectOption` — which already carried both — gains nothing.
329
+ **`ds-autocomplete` now acts on a `disabled` option**, where it used to ignore the flag entirely: the
330
+ row carries `.is-disabled` and `aria-disabled`, refuses the click, and **the arrow keys step OVER it** —
331
+ including when the panel opens or the filter re-runs, so the active row is never one that Enter would
332
+ ignore. Behaviourally this matches `ds-select` and `ds-multi-select`, which have honoured it all along.
333
+ The row greys out too: measured in the open panel it comes out `rgb(155,163,180)` against
334
+ `rgb(22,26,40)` for an enabled one, so the state is not subtle.
335
+
336
+ - **`maxVisibleChips="fit"` on `ds-autocomplete` and `ds-multi-select`** caps the chips by measured
337
+ width instead of a count: one line holds as many as the trigger is wide, the rest collapse into
338
+ "+N more", and the split is redone whenever the trigger resizes. **It is also the new default** — see
339
+ the Changed entry below for what that moves and how to opt out. Choose between the three by what has
340
+ to stay predictable: `"fit"` lets the count follow the width, a NUMBER adds an upper bound on top of
341
+ the same measuring, and `null` is the only mode that shows every chip and the only one that never
342
+ measures. **You can now delete** a `ResizeObserver`, container query or per-breakpoint
343
+ `maxVisibleChips` number written to keep a chip trigger on one line — and a `max-height` +
344
+ `overflow: hidden` cap, which clips a chip mid-glyph and hides the count instead of dropping whole
345
+ chips. The measurement is not free but it is bounded: it runs only where a chip row is rendered —
346
+ which on `ds-multi-select` means a non-empty selection, but on `ds-autocomplete` means any `multiple`
347
+ trigger, empty or not — widths are cached per selection, and a geometry guard skips the re-measure
348
+ when neither width nor count moved.
349
+
350
+ - **`--ds-dropdown-panel-radius` — the lever 0.9.0's notes promised and never shipped.** That release
351
+ said an 8px select panel could be restored with `--ds-select-panel-radius`; no such token existed,
352
+ so the advice did nothing (the note has been corrected). This is the honest version, and it is
353
+ panel-level rather than per-component on purpose: retuning one picker's corner is not a thing to
354
+ encourage, while `--ds-radius-panel` moves modal, popover and date picker too. Defaults to that
355
+ role, so an untouched theme is unchanged, and it is forwarded onto the portaled pane — scope it with
356
+ a selector (`ds-select { --ds-dropdown-panel-radius: 8px }`) and only that panel moves. Verified
357
+ in a browser: one select at 3px, the other 34 on the 12px role. **Two panels read it, not three:**
358
+ `ds-select` and `ds-autocomplete`. `ds-multi-select`'s panel hardcodes `--radius-md` and is reached by
359
+ neither this token nor `--ds-radius-panel` — so only `--radius-md` itself moves it, and that moves
360
+ buttons, fields and tooltips with it.
361
+
362
+ - **Two migration manifests were corrected and three more were written**, so re-read them if you
363
+ subclass a DS component or upgrade across a version you skipped. `0.9.0`'s named no member collisions
364
+ at all — the note that now fills the gap lists **fifteen** classes whose member surface had grown, plus
365
+ four newly exported symbols that can shadow a local name — and `0.8.0`'s missed the private
366
+ `saveLayout` that actually stopped a build (`restoreLayout` went in beside it). **`0.6.1`, `0.6.2` and
367
+ `0.9.1` had no manifest at all**, and now every released version carries one: the auto-update planner
368
+ prefers a manifest and otherwise re-derives the plan from changelog prose on every run, which is what
369
+ an upgrade crossing those three used to do. `0.6.2`'s turned out to hold something actionable that no
370
+ machine-readable file had ever stated — a `ds-bottom-sheet` with nothing focusable inside it stopped
371
+ closing on Escape — which is exactly the kind a "just fixes" patch is assumed not to have.
372
+
373
+ - **`ds-avatar` takes a `name`, and the rule that reads it is replaceable.** One binding now does both
374
+ jobs a caller used to wire by hand: it derives the initials and seeds the tint. **The tint hashes the
375
+ NAME, not the initials** — two letters collide wholesale, so Alice Brown and Adam Bell used to share a
376
+ colour whenever a caller seeded from the letters. Swap the rule with `[initialsRule]` on one table or
377
+ `provideDsConfig({ avatar: { initials } })` for the app; an explicit `initials` still wins over both,
378
+ and `seed` still overrides the hash input when identity is an id or an email. The built-in rule is
379
+ `initialsFrom` — first two words — and the docs now say plainly that it is a Latin-script guess: how a
380
+ name is read is a cultural convention, not a UI decision, so the library takes yours rather than
381
+ shipping modes it cannot get right. **You can now delete** the `[seed]`/`[initials]` pair you compute
382
+ beside every avatar. `name` feeds the initials and the tint and nothing else: the avatar stays
383
+ `aria-hidden` unless you bind `ariaLabel`, so a `[name]` on its own leaves it silent to a screen
384
+ reader.
385
+
386
+ - **`DS_CONFIG.brand` — set the brand once instead of on every screen.**
387
+ `provideDsConfig({ brand: { name, logoLetter, tagline } })` answers `ds-sidenav` and `ds-auth-layout`
388
+ together — `name` and `logoLetter` on both, `tagline` on the auth layout, which is the only one of the
389
+ two with a strapline row. They ask the same questions and an app answers them identically, so six auth
390
+ screens were binding the same values six times. Each field resolves on its own, so a screen can
391
+ override the name without restating the letter — and `brand=""` drops the wordmark while a
392
+ `logoLetter` in the config
393
+ still paints its tile, so clear both to give a screen no brand row. **The two-tone wordmark stays
394
+ projected:** there is deliberately no accent in the config, because a string would have to be split
395
+ back apart to find one and that is the guessing `[dsBrandName]` exists to avoid. The projection wins
396
+ over both. **You can now delete** an adapter component that existed to inject a brand object and fan
397
+ it out into four bindings.
398
+
399
+ - **`rememberLastTab` on `ds-tabs` — reopen the tab a screen was last left on.** Off by default, and
400
+ inert without `urlParam` (and outside a routed context), because the parameter supplies the names: a
401
+ remembered INDEX would repoint the moment a tab is inserted. Stored under
402
+ `ds-tabs:{path}:{param}` in `sessionStorage`, so two strips on one path stay apart and tomorrow is a
403
+ fresh visit. **A link always wins** — the query parameter is read first, and memory answers whenever
404
+ it is absent, which is any arrival at the bare path, the deliberately-empty first tab included.
405
+ Tab one is remembered too, or going back to it would be forgotten. Only a user's own click or
406
+ keypress writes the memory; a host driving `selectedIndex` writes neither it nor the URL, as before.
407
+ **You can now delete** a helper that kept the last tab per path.
408
+
409
+ - **`DS_CONFIG.modal.dividers` — turn modal rules off for a whole app in one declaration.**
410
+ `provideDsConfig({ modal: { dividers: false } })` instead of a binding on every `<ds-modal>`, with an
411
+ explicit `[dividers]` on an element still winning. It also reaches where an attribute cannot: a
412
+ `ModalService` confirm renders a `ds-modal-card` with no `ds-modal` ancestor to bind, and the injected
413
+ default gets there (verified — the confirm's card comes out `no-dividers`). One limit on that: the
414
+ service is `providedIn: 'root'` and builds the confirm's injector from the root, so only a ROOT-level
415
+ `provideDsConfig` reaches it — a route- or component-scoped one does not. **You can now delete** a
416
+ `dividers` binding repeated per modal, and any `!important` reach into `.modal > .header|.body|.actions`
417
+ written because the token swap left the padding stacked. Note `[dividers]` now reads
418
+ `boolean | null` internally so "unset" is distinguishable from `false`. A literal `[dividers]="false"`
419
+ is unaffected; a binding that can evaluate to `null`/`undefined` is not, since that used to mean
420
+ "off" and now means "unset", which resolves to `true`.
421
+
422
+ - **`triggerSize` on `ds-filter-bar`, to match its "+N" trigger to the filters beside it.** Default
423
+ `md`, so nothing moves. The bar cannot work this out for itself — `dsFilterItem` carries a width and
424
+ the filters are arbitrary projected controls whose own `size` it never sees — so a row of `sm`
425
+ filters got an `md` trigger standing 8px taller under `align-items: flex-end`. **You can now delete**
426
+ `--ds-button-height` / `-font-size` / `-pad-x` pinned under `ds-filter-bar` to force this.
427
+
428
+ - **`[debounce]` on `ds-search`, so a list screen stops wrapping it.** `[debounce]="200"` holds each
429
+ keystroke back before `value` updates; the default `0` writes on every keystroke exactly as before,
430
+ and what the input SHOWS is never delayed at any setting. **Three things skip the timer on purpose**,
431
+ which is what a wrapped `debounceTime` usually gets wrong: emptying the field, the Clear button, and
432
+ Enter — the last flushing a keystroke still in flight so `submitted` carries what the reader can see.
433
+ Emptying also cancels a pending keystroke, so type-then-clear cannot land the stale query a moment
434
+ later. A host writing `value` mid-typing wins and cancels the timer. The Clear button follows what is
435
+ ON SCREEN rather than the debounced `value`, so at a non-zero delay it appears on the first keystroke
436
+ instead of after the wait. **You can now delete** a debounce wrapper around this control.
437
+
438
+ - **`--ds-stat-value-white-space` and `--ds-stat-value-font-size`, for a tile whose value is a phrase.**
439
+ The value line is `nowrap`, which is right for the number `ds-stat` was built around and wrong for a
440
+ date range or a status — so a text value spilled straight out past the tile's edge, and the way out
441
+ was the projected
442
+ `[dsStatValue]` slot plus two declarations undoing the component. Both default to today's values, so
443
+ nothing moves until you set them. **The size hook works at every `size`**, not just the default: the
444
+ `sm`/`lg` variants now move the same token instead of restating `font-size`, which used to out-rank
445
+ any override. **You can now delete** a `[dsStatValue]` span that exists only to set `white-space`
446
+ or a smaller size.
447
+
448
+ - **The global `.link` class now works on a `<button>`, which it has been documented as doing all
449
+ along.** `ds-button` warns `variant="text"` off in favour of `.link`, and an action that is not
450
+ navigation has to be a button — but the class was written for anchors and reset none of the browser's
451
+ button chrome, so everyone who took that advice wrote the same rule by hand. Measured against an
452
+ `<a class="link">` in the same paragraph, a button came out at 13.3px inside 14px text, with
453
+ `line-height: normal`, `1px 6px` of padding and an outset border; all of that is gone. **You can now
454
+ delete** a local `.link-button`, or the padding/border/background/font reset in your overrides.
455
+ `font-weight` is untouched so `.quiet` and `.standalone` still decide it, and `display` stays
456
+ `inline-block` so a standalone button lays out predictably. The reset is scoped with `:where()`, so
457
+ it sits at `.link`'s own specificity and your rule still wins on source order.
458
+
459
+ ### Fixed
460
+
461
+ - **A card no longer eats the Enter and Space of everything inside it.** `ds-card` bound
462
+ `(keydown.enter)` and `(keydown.space)` on its `<article>` with an unconditional `preventDefault()`,
463
+ and keydown bubbles — so a card cancelled the DEFAULT ACTION of both keys for its whole subtree, on
464
+ every card, not only the `elevation="interactive"` ones that are supposed to answer a key at all.
465
+ Measured on a sign-in screen made of `ds-card` + `ds-input` + `ds-checkbox`, driven with real key
466
+ events: Enter in the password field sent **no request at all** — the form could only be submitted by
467
+ clicking the button — Space did not toggle the "Remember me" box, Enter on a link inside the card did
468
+ not follow it, and a space could not be typed into a text field, so `John` + Space + `Smith` came out
469
+ as `JohnSmith`. Both handlers now run only on an interactive card, and only when the key landed on
470
+ the card itself rather than on something projected into it.
471
+
472
+ - **A table row's own controls answer the keyboard again.** `onRowKey` consumed Enter and Space for
473
+ every target inside an interactive `<tr>`, so the row's select checkbox could not be toggled with
474
+ Space and a cell's button could not be activated with Enter — the row-actions button carries a
475
+ `stopPropagation` workaround for exactly this. The row now opens only on a key pressed on the row
476
+ itself.
477
+
478
+ - **`ds-menu`'s type-ahead no longer swallows a projected field's typing.** Every printable character
479
+ reaching the menu jumped focus to an item and was consumed, which left a search or filter input
480
+ placed inside a menu impossible to type into. Keys from a text-entry control now pass through.
481
+
482
+ - **`ds-radio-group` no longer swallows Enter, so a form still submits from inside it.** The group
483
+ handled Enter as "select the focused radio" and prevented its default, which is neither native radio
484
+ behaviour nor what APG asks for — and inside a form it silently removed the keyboard's way to submit.
485
+ Enter is no longer handled; Space still selects. A text field sitting among the radios (the
486
+ `Other: ___` row) also keeps its own keys now.
487
+
488
+ - **A field in error no longer tints itself inside a `ds-field-group`, and an invalid group draws a red
489
+ border rather than a near-black one.** Two faults the field-state work left behind, both found by an
490
+ adversarial re-read of this changelog. The state rules are included at the end of each stylesheet and
491
+ score the same `(0,3,0)` as the `.in-group` block that strips a segment's chrome, so source order
492
+ alone decided: measured, the border stayed 0 (`border: 0` zeroes the width, so a later `border-color`
493
+ paints nothing) but the FILL came back, tinting one segment pink mid-row — the exact thing a group
494
+ exists to prevent. All three states now carry `:not(.in-group)`. Separately `ds-field-group` itself
495
+ read `var(--border-error)`, **a token declared nowhere**, so `[invalid]` fell through to
496
+ `currentColor`; it now resolves through a new `--ds-field-group-border-color-error`, defaulting to
497
+ `--error`, matching every other colour the component publishes.
498
+
499
+ - **`ds-timeline-item`'s marker sits level with its heading.** The node is 28px and the heading's line
500
+ box is 19.6px, and both started at the row's top edge — so the circle's centre sat **4.2px** below the
501
+ heading's and read as floating between the heading and the time beneath it. The node column is now
502
+ pulled up by half the difference, computed from the same `--font-size-s` / `--line-height-base` pair
503
+ the heading uses, so retuning the heading moves the marker with it. Measured across all ten items on
504
+ the docs page: every one now aligns to 0.
505
+
506
+ - **The collapsed `ds-sidenav` stopped squashing its items into a lozenge.** Every item in the rail
507
+ measured **20×40** where the rail gives it 40×40, so the active item's blue tile rendered as a
508
+ vertical bean with the orange rail bar down one side of it — and the alert dot, positioned from the
509
+ item's corner, landed above the glyph's middle instead of beside it. One cause for both:
510
+ `:host(.collapsed)` set `align-items: center`, which stops a flex child stretching, so each projected
511
+ `ds-sidenav-item` and `ds-sidenav-group` shrank to its 20px icon and the `width: 100%` on `.item`
512
+ resolved against *that*. The rail no longer centres its children — `.brand` already centres its own
513
+ contents, which is all that line was there for. Items now measure 39×40 (the rail's 64px less its
514
+ padding and border). The dot is anchored to the GLYPH's top-right corner rather than the tile's, via a
515
+ new `--sidenav-collapsed-glyph`, so it stays put if the rail width ever changes.
516
+ **ACTION** only if you relied on a projected child of a collapsed `ds-sidenav` sizing to its own
517
+ content — a custom element in the rail now stretches to the full 40px column.
518
+
519
+ - **`ds-time-picker`: the time you just picked stopped looking selected.** Hovering a chosen hour or
520
+ minute painted it with the neutral hover fill while it kept the selected white text — white on pale
521
+ grey, unreadable. Pure specificity: `.opt:hover:not(:disabled)` scores (0,3,0) and out-ranked
522
+ `.opt.is-selected` at (0,2,0). It read as a *touch* bug because a touch screen leaves `:hover` on the
523
+ last-tapped cell until something else is tapped, so the minute you had just chosen sat greyed out
524
+ until you looked away. Hover now skips a selected option; an unselected one is unchanged.
525
+
526
+ - **`ds-time-picker` in single mode closes itself, and `Done` is gone from it.** Choosing an hour and a
527
+ minute *is* the completion, so a confirm button underneath it was a second way to say the same thing.
528
+ The panel now closes once both columns have been used **in that opening** — re-opening a field that
529
+ already holds a value asks for both again rather than closing on the first tap, and changing only the
530
+ hour leaves the panel up. `Now` takes the full width of the footer beneath the columns, so it reads as
531
+ one action across the panel rather than a button hugging the right edge. **`mode="range"` is
532
+ untouched**: it keeps `Done`, and picking both halves there does not close anything, because the end
533
+ time still has to be chosen. Internally `ds-time-columns` gained a `partPicked` output — `timeChange`
534
+ carries a whole time and cannot distinguish "chose the same hour again" from "never touched hours".
535
+
536
+ - **`ds-password-rules`: a pending rule is an empty circle, and a rule fades as it passes.** The
537
+ pending glyph was `minus`, which at 14px beside 12px text read as a stray em-dash rather than an
538
+ unticked box; it is now `circle` — the empty counterpart of the check it turns into. A `circle` icon
539
+ joins the set (86 now), available to anything else that wants one. The row also transitions its
540
+ colour over `--duration-base` instead of snapping, because the reader is looking at the field they
541
+ are typing in, so an instant change says a rule flipped but not which one. No reduced-motion guard
542
+ is needed: the duration tokens are already zeroed globally. Row spacing is unchanged at
543
+ `--space-1` — deliberately, after comparing both.
544
+
545
+ - **`ds-password-rules` printed its screen-reader text on screen.** Every row rendered a visible
546
+ "— not met" / "— not yet met" beside the label, because the component used `.visually-hidden` and
547
+ **nothing anywhere defined it**: three other components (`ds-table`, `ds-date-picker`, `ds-stepper`)
548
+ each declared their own byte-identical copy, so the class only ever worked by accident of being in
549
+ the right component. The leaked text is also what made the two-column layout look broken on a phone —
550
+ it ate the column width and wrapped every label onto two lines, which read as a spacing failure and
551
+ was not one. One `visually-hidden` mixin now backs all four. **No API change**; a checklist just
552
+ stops saying "not met" four times in a row. The host also takes `position: relative`, to contain those
553
+ now-hidden rows — so if you absolutely position anything inside `ds-password-rules`, it anchors to the
554
+ component rather than to whatever ancestor used to hold it.
555
+
556
+ - **`ds-otp-input` no longer shows two digits in one cell.** Typing a second character into a cell that
557
+ already held one left `12` sitting in that box while the model had already split it correctly across
558
+ two cells — so what the user read and what the form held disagreed. The distribution was never the
559
+ bug: `[value]` was. Angular diffs a binding against the value *it* last wrote, and the typed-in cell's
560
+ signal is unchanged by the split (`'1'` before and after), so the DOM write was skipped and the
561
+ browser's own `"12"` survived. That branch now writes the element itself, like the other two already
562
+ did. Pasting a full code is unaffected — it cancels the native input, so nothing ever reaches the box.
563
+
564
+ - **A `warning` now shows on the control, and one rule drives all nine bordered fields.** `warning="…"`
565
+ coloured the message and left the box untouched — which is what a phone screenshot caught. On 0.9.1
566
+ six stylesheets styled `.warning` and not one belonged to a bordered field: `ds-field-message`, the
567
+ message this describes, plus `ds-alert`, `ds-banner`, `ds-progress`, `ds-sidenav-footer` and
568
+ `ds-toast`. It now takes a `--warning-text` border. Not `--warning`:
569
+ that is `--yellow-400`, **1.47:1** against a white field where WCAG 1.4.11 asks 3:1 for a component
570
+ boundary, so the obvious mirror of the error recipe would have drawn a border nobody can see;
571
+ `--warning-text` is the first rung of the ramp that clears it, at 4.59:1. Border only, deliberately —
572
+ an error blocks submission and keeps border + tinted fill, so it still reads as the loudest thing in
573
+ a form holding all three.
574
+
575
+ The nine had drifted badly and are now one shared `field-state` mixin keyed off a `data-field-state`
576
+ attribute: five painted the error box as border + fill and four as a border alone;
577
+ **`ds-autocomplete` reached for a `--border-error` token that does not exist**, so an invalid
578
+ autocomplete drew a *near-black* border (`rgb(22,26,40)` — `border-color` falling back to
579
+ `currentColor`) instead of a red one; and the same state was spelled four ways — `.error` on the
580
+ wrapper in `ds-input` and `ds-textarea`, `.error` on the box itself in `ds-select`, `ds-date-picker`,
581
+ `ds-time-picker` and `ds-autocomplete`, `.invalid` on the box in `ds-multi-select`, and
582
+ `:host(.invalid)` in `ds-color-picker` and `ds-otp-input`. **What changes visually:**
583
+ `ds-autocomplete` finally shows red, and `ds-multi-select`, `ds-color-picker` and `ds-otp-input` gain
584
+ the tinted fill their five siblings already had. A sixth thing moves that is easy to miss:
585
+ `ds-autocomplete`, `ds-multi-select`, `ds-date-picker`, `ds-time-picker` and `ds-color-picker` had no
586
+ error FOCUS ring at all and kept their ordinary accent one while invalid — the shared mixin gives all
587
+ nine `--ds-field-focus-shadow-error`. Only `ds-input`, `ds-select`, `ds-textarea` and `ds-otp-input`
588
+ carry a per-component override for it, so retuning the other five means moving the family token,
589
+ which moves those four with it. One precedence moved with the shared rule: the ladder
590
+ is now error → warning → success, so a `ds-input` carrying both `[success]="true"` and a `warning`
591
+ string paints the warning, where it used to paint the success and ignore the warning entirely. Every
592
+ per-component override token that shipped in 0.7.x (`--ds-input-bg-color-error`,
593
+ `--ds-select-focus-shadow-error`, …) still wins — the mixin takes them as arguments rather than
594
+ replacing them with the family token. The `.error` class is **still emitted** — on the wrapper by
595
+ `ds-input` and `ds-textarea`, on the box by `ds-select`, `ds-date-picker`, `ds-time-picker` and
596
+ `ds-autocomplete` — and `.success` by `ds-input`, the only control with a success state. A selector of
597
+ yours still matches, but the library no longer paints through them, so a rule written to stack on top
598
+ of the DS one now stands alone.
599
+
600
+ **DELETE-LIST:** the dead `.invalid` host class on `ds-color-picker` and `ds-otp-input` is gone, since
601
+ nothing styled it any more; if you selected it from outside, use `[data-field-state="error"]`.
602
+ `ds-multi-select` loses the same class from its internal `.field-trigger` — it was never on the host,
603
+ so `ds-multi-select.invalid` never matched from outside anyway. **The attribute is on those nine and
604
+ only those nine.** `ds-checkbox`, `ds-radio`, `ds-radio-group` and `ds-switch` do not carry it: they
605
+ have no bordered box, they signal through `aria-invalid` and `ds-field-message`, and their paint is
606
+ untouched. Nor does `ds-slider`, which keeps its own `:host(.invalid)` binding and still colours its
607
+ fill and thumb `--error` — so a selector written against `[data-field-state]` will not reach any of
608
+ the five.
609
+
610
+ - **Binding `error` now tints the control, not just the message.** Every field that takes an `error`
611
+ string — all 14, via the shared `resolveFieldMessage` — put a red message under a box that stayed
612
+ neutral grey, because the invalid state was computed from `[invalid]` and from validator `errors` but
613
+ not from an explicit `error`. That contradicted the input's own contract ("binding it says the
614
+ consumer has already decided this field is in error") and it is what a phone screenshot caught: a red
615
+ line under an untouched field. `error="…"` now drives the red border, the tinted fill, the error focus
616
+ ring and `aria-invalid`, exactly as `[invalid]="true"` always did — the border and fill on the nine
617
+ bordered fields, and `aria-invalid` plus the message state on the other five, which have no box to
618
+ paint. **If you were binding both**
619
+ (`[invalid]="hasError" error="…"`) nothing changes. **If you were binding `error` alone and relying on
620
+ the box staying neutral**, that is now the error box — drop to `warning` for an advisory message, or
621
+ bind `hint` and leave `error` empty.
622
+
623
+ - **A field inside `ds-field-group` no longer pins its value to the top of the row.** Found by Stas
624
+ testing the docs on a phone, and it affected **every segment except `ds-button`**: the value sat
625
+ against the top edge instead of the vertical centre — **5px out at `sm`, 9px at `md`, 13px at `lg`**,
626
+ measured in one sitting against the 21.6px content height the box fell back to on the 32/40/48 row
627
+ scale. The group stretches its segments to one row height, but a segment's host is a block, so the
628
+ `height: 100%` each one set on its bordered box resolved against `auto` — the box kept that content
629
+ height, and the value with it. `ds-button`
630
+ escaped the defect only because its host already declared `display: inline-flex`; its block was
631
+ rewritten with the rest all the same, to `flex: 1; height: auto`. Every segment now centres at all
632
+ three sizes. The six `.in-group` blocks — identical declaration for declaration apart from the box's
633
+ class name, which is
634
+ why the defect sat in five of them at once — are now one `field-group-segment` mixin, so the next
635
+ change to segment chrome cannot land in some of them and miss the rest.
636
+
637
+ - **`components.md` had drifted, and it ships inside the package.** The file `tools/build-schematics.mjs`
638
+ copies into `dist/ds/` for a consumer's own agents to read was still documenting the API and token names
639
+ of two renames ago: **46 of the 92 `--ds-*` tokens it named no longer exist** (`--ds-button-fg`,
640
+ `--ds-sidenav-item-bg-hover`, the whole `--ds-pagehdr-*` family, …), `ds-button` was described with its
641
+ pre-2026-06 single `variant` axis (`'primary' | 'secondary' | 'tertiary' | …`) instead of
642
+ `color` × `variant`, and the topbar section documented a `--ds-topbar-margin-b` that **has never
643
+ existed** — telling readers the topbar owns the gap below it and that pages must not add their own.
644
+ Every name in the body is now the shipped one, verified against `src/lib` rather than assumed, and the
645
+ two paragraphs are rewritten from the components' actual API. The preamble of its `## Components`
646
+ section still carries the older warning that the `ds-button` section below is out of date — which now
647
+ contradicts the corrected section, and is itself wrong twice over: it calls `--ds-button-bg-color` a
648
+ dead pre-rename name when that is the shipped token this very release tells you to set. Being cleared
649
+ separately.
650
+
651
+ - **An outlined button no longer paints its own white background.** `ds-button variant="outline"`,
652
+ `ds-icon-button variant="outlined"` and the outlined half of `ds-split-button` defaulted their fill to
653
+ `--surface-default`, which is invisible on a white page and wrong everywhere else: on a tinted card, a
654
+ table toolbar or an alert the button read as a white sticker laid over the surface instead of a control
655
+ drawn on it. The fill is now `transparent`, which is what Material 3 and Bootstrap both specify for this
656
+ variant. Hover still tints — `--c-subtitle` on `ds-button`, `--surface-secondary` on the other two, all
657
+ opaque. On `ds-button` and `ds-icon-button` the disabled state stays transparent as well, because mixing
658
+ a transparent fill toward the disabled surface would have produced a 72%-alpha box, so the border and
659
+ label carry it, as they do on ghost; a disabled `ds-split-button` is the exception and still paints
660
+ `--surface-muted`, which this change did not reach. **If a button of yours sits on a photo, a gradient
661
+ or any busy ground and relied on that white fill**, set it back per instance:
662
+ `--ds-button-bg-color: var(--surface-default)`, or `--ds-icon-button-bg-color` /
663
+ `--ds-split-button-outline-bg-color` for the other two.
664
+
665
+ - **Everything focusable now gets the DS focus ring, not the browser's black outline.** The library
666
+ styled focus on the elements that remembered to — 42 of 81 component stylesheets, and only 35 of those
667
+ with `:focus-visible` — and everything else
668
+ fell through to `outline: auto`, which Chrome paints black: the close button inside `ds-alert`,
669
+ `ds-banner` and `ds-drawer`, a link in projected content, and any plain `<button>` an app writes itself.
670
+ One base rule now covers them: `:where(:focus-visible)` sets `outline: none` and resolves
671
+ `--ds-control-focus-shadow`, written at **zero specificity** so every component rule and every app rule
672
+ still out-ranks it. The `data-focus` switch is unaffected — all five modes (default, `ring`, `field`,
673
+ `edge`, `outline`) behave as before, the last replacing the shadow with a real outline at a higher
674
+ specificity. **If you deliberately suppressed a focus ring** by leaving an element unstyled, it will now
675
+ show one; give that element `box-shadow: none` explicitly. Note the rule also drops the outline such an
676
+ element used to get from the browser, so if an `outline` was what you wanted there, declare it yourself.
677
+
678
+ - **A disabled `ds-button` / `ds-icon-button` now reads as disabled on a COLOURED accent.** One label recipe
679
+ for every variant — `oklch(from accent 70% 0.035 h)` — replacing the 58% default, and on `ds-button` the
680
+ 52% overrides that tonal and outline carried (`ds-icon-button` only ever had the base). Those sat right
681
+ where a saturated accent already is, so the disabled label
682
+ differed from the live one by **1.16–1.60:1** on four of the five variants; only `color="neutral"` was
683
+ safe, because its enabled label is nearly black. Now **2.40–2.71:1** on the coloured variants that
684
+ were failing, and higher still on `neutral`, which reaches 6.46:1.
685
+ Measured, not guessed: the matrix is in `presentations/disabled-recipe.html` — 5 variants x 3 colours
686
+ x rest/hover/active/disabled in light, and a narrower primary-only sweep in dark. **The trade,
687
+ stated:** the disabled label is fainter
688
+ against its own background than it was — 2.1–2.7:1 where the old recipe gave 3.4–5.2 — which is the point,
689
+ and disabled text is exempt from the WCAG contrast floor. **Override per component if you disagree:**
690
+ `--ds-button-label-color-disabled` / `--ds-icon-button-icon-color-disabled` still win. **And delete this
691
+ if you have it:** an override retuning `--opacity-muted` on `ds-button` or `ds-icon-button` never did
692
+ anything — neither component reads that token.
693
+
694
+ - **A disabled control was dimmed twice, and had nearly disappeared.** Five places each set
695
+ `--text-disabled` AND `opacity: 0.6` on top of it: the closed triggers of `ds-select`,
696
+ `ds-date-picker` and `ds-time-picker`, plus two genuine rows — `ds-menu-item` and the shared dropdown
697
+ option. Measured in the light theme: that lands at **1.68:1** against the
698
+ surface, where the colour alone gives **2.53:1** and an enabled control sits at 17.32:1 — so the
699
+ state was
700
+ never in danger of being missed, the text was. The opacity is gone from those five; the colour carries
701
+ the state, which is what it was chosen for. Where opacity IS the only signal — a disabled calendar day,
702
+ `ds-toggle-button` — it now reads `--opacity-disabled` (0.5) instead of `--opacity-muted` (0.6), the
703
+ token named for the job. **If you retuned `--opacity-muted` to compensate**, drop it: it is a
704
+ de-emphasis level for hints and secondary text, and it no longer carries any disabled state.
705
+
706
+ - **`ds-table` lost the sort indicator when a host answered a cleared sort with its own default.** Third
707
+ click clears a column, the host replies with its default sort, the rows come back ordered by it — and
708
+ every header said `aria-sort="none"`. Not the table's arithmetic: Angular writes an input only when the
709
+ bound EXPRESSION changed, diffing against the last value it wrote, which an internal `.set()` never
710
+ updates. So the half of the default that equalled what was last bound was skipped — a host defaulting
711
+ to `owner`/`desc` had its `desc` dropped and stayed at `null`, while a host defaulting to `asc` was
712
+ restored. That is exactly why the same three clicks worked on one screen and not another. The marked
713
+ state is now a `linkedSignal` over BOTH inputs, so a change in either re-seeds the pair and the
714
+ direction arrives with the key. **Verified against both reported screens** by replaying them through
715
+ Angular's real signal graph with the binding diff modelled: the old shape fails the first and passes
716
+ the second, the new one passes both. A table that binds neither input is unaffected.
717
+
718
+ - **`ds-table` painted its empty state while the first read was still in flight.** Only
719
+ `loading="skeleton"` suppressed the empty row, and `overlay` is the mode a list uses on every read
720
+ after the first — so a zero-row reload showed "No records match your filters" or a whole
721
+ `ds-empty-state` CTA underneath the spinner. The empty row is a claim about the data ("there are
722
+ none"), so it is now withheld whenever `loading` is anything but `false`. **You can delete the guard
723
+ you wrote for this** — and check the ones you did not write: of one consumer's 14 overlay-loading
724
+ tables, 9 were shipping the sentence during a load, because branching failure-vs-empty looks complete
725
+ while saying nothing about loading. Note this closes the door on the `[empty]` slot during a load
726
+ rather than opening another one: there is no `loading` value that renders it, so copy you want a
727
+ reader to see WHILE a load runs has to come from your own page, above or around the table.
728
+
729
+ - **Releasing the mouse on the backdrop closed `ds-modal` and `ds-bottom-sheet`, even when the press
730
+ started inside the dialog.** Select a sentence in the body and let go past the edge, or overshoot with
731
+ a slider, and the dialog vanished — reported by a consumer as "not a click, the mouse-UP". The cause is
732
+ in the DOM shape: those backdrops WRAP their dialog, and a browser fires `click` at the nearest common
733
+ ancestor of the press and the release, which for that gesture IS the backdrop — so the
734
+ `target === currentTarget` guard let it through. Dismissal now tracks the press and fires only when
735
+ both the press and the release land on the backdrop itself, which is how the CDK has always guarded
736
+ its own outside-click. `ds-drawer` and the `ModalService` confirm were never affected (their scrim is
737
+ a SIBLING of the panel, so the common ancestor is a third element) and are unchanged — all four were
738
+ re-checked. **If a test of yours dismisses a dialog with `backdropEl.click()`**, it now needs
739
+ `pointerdown` + `pointerup` on the backdrop, both with the primary button; a real mouse or touch has
740
+ always sent both. One side effect on `ds-bottom-sheet`: the sheet no longer stops click propagation
741
+ (the guard it needed for the old `click` listener is gone with it), so a click inside the sheet now
742
+ reaches an ancestor handler of yours, the sheet being rendered inline rather than portaled.
743
+
744
+ - **`ds-filter-bar` clipped the focus ring of its first and last inline filter.** The measured row was
745
+ `overflow: hidden` and each slot is sized to exactly the width `dsFilterItem` declared, so the row's
746
+ box starts flush with the first control — and the focus ring is an OUTSET shadow (3px on fields, 4px
747
+ on controls). Interior filters were fine, living in the 12px gap; the outer two had their ring cut off
748
+ as you tabbed onto them. Now `overflow: clip` with a 4px clip margin: the shadow bleeds, `clip` still
749
+ forbids scrolling, and the width the collapse arithmetic reads is unchanged. Not fixable from an app —
750
+ the row is internal, and padding it from outside would change that measured width.
751
+
752
+ - **A single-select `ds-autocomplete` kept the text you typed to filter, and since 0.9.0 kept it
753
+ INVISIBLY.** The filter input collapses to zero width once a chosen label is on show, so leftover
754
+ query text disappeared from view while it went on narrowing the list: reopening a settled picker
755
+ offered one option out of four, with nothing on screen to say why and no way to clear it. Before
756
+ 0.9.0 the same text simply sat beside the label instead. It is now cleared whenever the panel closes
757
+ over a displayed label — which covers both picking an option and dismissing after typing over a
758
+ settled value. Multi-select is deliberately unchanged: its input never collapses, so the text stays
759
+ visible and useful for the next pick. Nothing to do on your side; `query` was component-private, so
760
+ this was not fixable from an app.
761
+
762
+ - **Four defects in `ds-sidenav`'s brand row.** A `logoLetter` set without a `brand` string rendered
763
+ **nothing at all** — if you worked around that by passing a dummy `brand`, you can drop it now. A
764
+ brand starting with an emoji or a rarer CJK glyph painted as tofu in the letter-mark. The mark had
765
+ no `aria-hidden`, so a screen reader announced the glyph and then the wordmark ("A MyAdmin"). And a
766
+ logo projected into `[brand]` had no size constraint, so an oversized one set the rail's height —
767
+ `--ds-sidenav-brand-max-height` now caps it, defaulting to `none` so nothing moves until you set it.
768
+
769
+ - **`env(safe-area-inset-*)` needs one line from YOUR app, and nothing here said so.** The insets read
770
+ 0 until the host document sets `viewport-fit=cover` in its viewport meta — a notched device alone is
771
+ not enough, and a library cannot set it for you. So `ds-auth-layout`'s notch padding and
772
+ `ds-bottom-sheet`'s home-indicator padding have been inert in every app that never added it. Both
773
+ component pages now say so and show the tag, and the docs site sets it, so the behaviour can be seen
774
+ rather than described. **One real fix rode along:** `ds-bottom-sheet` was calling
775
+ `env(safe-area-inset-bottom)` with no fallback inside a `padding` shorthand, and an engine that does
776
+ not know the variable throws out the whole declaration — taking the sheet's horizontal padding with
777
+ it. It now reads `env(safe-area-inset-bottom, 0px)`, which is what `ds-auth-layout` — the only other
778
+ component using these insets — already did on all four of its.
779
+
780
+ ### Changed
781
+
782
+ - **The seven inputs every field shares are declared once, in `DsFieldBase`.** `hint`, `error`,
783
+ `warning`, `errors`, `invalid`, `touched` and `reserveMessage` — plus the `reserveMessageResolved`
784
+ helper — were hand-copied into all fourteen fields, 98 declarations. **Nothing about the public API
785
+ changed:** every input keeps its name, type, default and documentation, and the docs API tables are
786
+ byte-identical apart from the eight new `success` rows (verified by diffing the generated
787
+ reference: 0 lost, 8 gained).
788
+ **Why it is worth knowing anyway:** the duplication had already drifted twice — `success` was
789
+ missing from eight fields, and `ds-date-picker` had quietly moved to a different `invalid` source.
790
+ A new state added to the base now reaches every field at once.
791
+ `DsControlValueAccessor` and `DsBooleanControlValueAccessor` extend the base (so single inheritance
792
+ still holds for the thirteen fields that already extended one of them) and `ds-radio` extends it
793
+ directly. Only the UNIVERSAL inputs moved: `success`, `label`, `required` and `optional` stay on the
794
+ components, because they exist on 9, 9, 7 and 5 fields respectively and lifting them would hand a
795
+ checkbox a `success` or a slider a `label` it never had.
796
+
797
+ - **`ds-icon-button` with `variant="standard"` mixes its disabled glyph toward the surface too.**
798
+ New default on that variant only: `color-mix(in srgb, <icon colour> 40%, var(--surface-default))`.
799
+ Same defect as ghost, and it matters more here — there is no word to read, just a 16px glyph.
800
+ Measured: light **4.17 → 5.19:1**, dark **1.46 → 3.58:1**. Light was already healthy, because the
801
+ enabled glyph is `--text-secondary` at OKLCH L 35% and the old recipe moved it to L 70%; **dark was
802
+ the break**, where the enabled glyph is L 80% and a disabled one pinned at L 70% left ten points
803
+ and a 1.46:1 gap.
804
+ **Scoped to `standard` deliberately, and this is the trap worth knowing:** the resolver is declared
805
+ once on the base and no variant overrode it, so touching the base would have hit `filled` too —
806
+ whose glyph is `--c-on`, i.e. white, and white mixed into a white surface does not move at all,
807
+ which would have put an invisible white glyph on a pale disabled fill. `filled`, `tonal` and
808
+ `outlined` keep the base recipe unchanged. Consumers who set
809
+ `--ds-icon-button-icon-color-disabled` are unaffected.
810
+
811
+ - **`avatarTint` switches from OKLCH to `hsl(H 50% 62%)` with white initials, always.** The seeded
812
+ tints read as dull, and every set of initials came out black. The auto-contrast was not the fault:
813
+ measured across the hue circle, the old `oklch(62% 0.14 h)` put every hue's luminance between 0.218
814
+ and 0.271 against a 0.179 black/white crossover, so black was genuinely the better pick every time
815
+ (5.4–6.5:1) — the switch was right, the lightness never let it switch. The new recipe is pinkmoon's
816
+ own, adopted deliberately: HSL lightness is not perceptually even, so yellows land far lighter than
817
+ blues, and that unevenness is exactly what makes the row read as bright.
818
+ **It is a knowing trade against contrast.** White on the lighter hues measures as little as
819
+ **1.65:1**, under the 4.5:1 WCAG AA asks of text — verified across the full 20–300° range, where all
820
+ 29 sampled hues fall short. Suitable where the avatar sits beside a name that is written out anyway;
821
+ where the initials are the only identifier, pass your own `lightness` / `foreground`.
822
+ **API:** the `chroma` option is gone, replaced by `saturation` (HSL %, default 50). `lightness` keeps
823
+ its name and its default of 62 but now means **HSL** lightness, so an existing numeric override will
824
+ land somewhere different — re-pick it rather than assuming it carried over. `foreground` now defaults
825
+ to `#fff` instead of auto-choosing; pass it explicitly to get anything else. `avoidHues` is unchanged.
826
+
827
+ - **`ds-input`'s `floating` label is removed, with no replacement.** It only ever did half the job: the
828
+ label was drawn inside a taller control, but it never moved out of the field on focus — the animation
829
+ from inside-the-box to above-it *is* the pattern, and without it an empty field reads as a
830
+ permanently mislabelled one rather than a floating-label one. Nobody had asked for it, and finishing
831
+ it is a real piece of motion work, so it goes. **Migration is deleting the attribute:**
832
+ `[floating]="true"` → nothing, and the label renders above the field like every other field in the
833
+ library. One thing to re-check when you do: a floating input stood at `--hit-touch` (48px at default
834
+ density) and now takes the ordinary size scale (`--field-h-md` → `--hit-cozy`, 40px at default
835
+ density), so a row whose height was built around one moves. Gone with it are the five internal
836
+ tokens only this variant read
837
+ (`--field-h-floating`, `--field-floating-label-fs`, `--field-floating-label-lh`,
838
+ `--field-floating-label-top`, `--field-floating-pad-t`) — overriding them is now inert.
839
+ `--ds-elevation-floating` is a different token for toasts and is untouched.
840
+
841
+ - **`ds-table`'s `sortKey` and `sortDir` are an `input` + `output` pair instead of a `model`** — the
842
+ mechanism behind the indicator fix above, because a `model` cannot tell a host's write apart from the
843
+ table's own. **Templates are unaffected:** `[(sortKey)]` still works (`sortKeyChange` and
844
+ `sortDirChange` are emitted), one-way `[sortKey]` with `(sortChange)` still works, and a static
845
+ `sortKey="project"` still works. What breaks is reaching into the component in TypeScript:
846
+ `table.sortKey.set(…)` no longer exists — nor `sortDir.set(…)` — so drive it from your own state and
847
+ bind it, which is what every server-paged list already does. Reading `table.sortKey()` still works and
848
+ now returns what the host bound; the marked state the header renders is internal. **One emission rule
849
+ moved with it:** both `…Change` outputs now fire on EVERY sort click, where a `model` suppressed the
850
+ emit when the value was unchanged. `[(sortKey)]` is unaffected — it rebinds the same value — but a
851
+ `(sortKeyChange)="reload()"` handler now also runs when only the DIRECTION changed. Listen to
852
+ `(sortChange)` instead, which carries both halves once.
853
+
854
+ - **`ds-modal` / `ds-modal-card`: `error: boolean` → `status: DsStatus`.** A boolean could not express
855
+ a warning or an info dialog, and the rest of the library settled on the `DsStatus` vocabulary in
856
+ 0.9.0. **Migration is one attribute:** `[error]="true"` → `status="error"`, and a computed
857
+ `[error]="isDangerous"` → `[status]="isDangerous && 'error'"`, since the input takes `false` and
858
+ `undefined` as "no status" — the longer `? 'error' : null` ternary works too. Unset stays neutral,
859
+ so a dialog that never used `error` looks identical. **What you gain beyond the rename:** a status now
860
+ drives the icon-badge tint, the title colour, a DEFAULT icon and the ARIA role — the same four
861
+ things `ds-alert`'s variant drives, so a status dialog and a status alert on one screen cannot
862
+ disagree. The default icon matters: a tint alone signals by colour only, which fails the
863
+ don't-rely-on-colour rule. `icon` still overrides the pick, and `error`/`warning` default the
864
+ dialog to `alertdialog`. `dialogRole` lives on `ds-modal-card` only — it is now
865
+ `'dialog' | 'alertdialog' | null`, defaulting to `null` rather than `'dialog'` so "unset" is
866
+ distinguishable from a deliberate `dialog`, and an explicit value still wins. `ds-modal` has no such
867
+ input, so `<ds-modal status="error">` becomes an `alertdialog` with nothing to override it. What
868
+ renders is unchanged unless you set a status. Three tokens come with the tints:
869
+ `--ds-modal-title-text-color-warning`, `-success` and `-info`, joining the `-error` that already
870
+ existed. **One thing rides along that no attribute of yours controls:** a
871
+ `ModalService.confirm({ tone: 'error' })` passes `status="error"` down internally, and a status now
872
+ supplies a default icon — so a destructive confirm that never set `options.icon` opens with a red
873
+ error badge where it previously had none. Pass `icon` to choose a different glyph; there is no input
874
+ that turns the badge back off.
875
+
876
+ - **Chips cap by width by default.** `maxVisibleChips` on `ds-autocomplete` and `ds-multi-select`
877
+ now defaults to `"fit"` instead of "show everything": the trigger keeps as many chips as it is
878
+ wide and collapses the rest into "+N more". A trigger that grew a line per selection was pushing
879
+ the rest of the form down the page. **To keep the old behaviour, bind `[maxVisibleChips]="null"`.**
880
+ Two more things moved with the default, and both are easy to miss. A trigger in any measuring mode is
881
+ now single-line (`flex-wrap: nowrap`), and its chips are held to one line each and clipped rather than
882
+ wrapped — so a chip carrying more than a short label gets squeezed, where `null` still wraps the row.
883
+ And **`[maxVisibleChips]="0"` has inverted**: it used to mean "show them all" and now means "show
884
+ none", putting the whole selection behind `+N more`. If you have a `0` anywhere, it was almost
885
+ certainly written for the old meaning; `null` is what that meant.
886
+
887
+ - **`size` is now `DsSize | null` on the fourteen controls that resolve `DS_CONFIG.control.size`**, so
888
+ an unbound `size` can fall through to it. `ds-drawer` and `ds-spinner` also take a `size` and are NOT
889
+ among them — theirs stays a non-null `DsSize` defaulting to `'md'`, so a grep for `DsSize` finds
890
+ sixteen controls and only fourteen changed. **Nothing changes in a template** — `size="sm"` binds as
891
+ before and an unbound control still renders `md`. **What changes is reading it back:** `size()` now
892
+ returns `null` when nothing was bound, where it used to return `'md'`. The same is true of
893
+ `ds-card`'s `elevation` and `padding`, `ds-icon`'s `size`, `ds-badge`'s `withDot`, `ds-table`'s
894
+ `bordered` and `stickyScrollbar`, `dividers` on both modal components, `dialogRole` on
895
+ `ds-modal-card`, `brand` and `logoLetter` on `ds-sidenav` and `ds-auth-layout`, and `tagline` on the
896
+ auth layout — all of them now answer `null` for "nothing bound"
897
+ rather than the old default. **There is no public accessor for the resolved value**, and this is the
898
+ awkward part of the release: each component holds a `sizeResolved()` (or `elevationResolved()`,
899
+ `withDotResolved()`, …) but declares it `protected`, so reading it from a host, a test or a
900
+ `viewChild` does not compile. `ds-field-group` is the single exception, where `sizeResolved` is
901
+ public. Until that is opened up, a host that needs to know what a control is rendering at should keep
902
+ the value it bound, or resolve `DS_CONFIG.control.size` itself — the fallbacks are unchanged, but
903
+ they are not all `false`: `'md'` for `size`, `'flat'` for `elevation`, `16` for the icon size,
904
+ `false` for `bordered` / `stickyScrollbar` / `withDot`, **`true`** for `dividers`, `''` for the brand
905
+ strings, and the card's own padding when `padding` is unset. **If you implement
906
+ `DsFieldGroupContext`**, its member is now
907
+ `sizeResolved` rather than `size`, for the same reason: a segment must not have to resolve the
908
+ group's size a second time.
909
+
910
+ - **`[data-density="ultra-compact"]` now compresses surface padding too** — `--ds-container-pad`,
911
+ the inner padding `ds-card` and `ds-stat` share, drops from 24px to 16px in that tier only.
912
+ Everything else is untouched: the default and the other three tiers still resolve to 24px. It was
913
+ the one padding the tier missed — it already compressed field padding and table cells, so a card
914
+ kept its full 24px frame while its own cells halved. **If you want the old 24px inside an
915
+ ultra-compact container**, re-declare `--ds-container-pad: var(--space-6)` on it. For a value
916
+ between the rungs — 20px, say — set that one token at `:root`; both components default to it, which
917
+ is why neither `ds-card` nor `ds-stat` should be pinned on its own.
918
+
919
+ - **`ds-tabs` `urlParam` now SLUGS a tab's label instead of using it raw.** A tab with no `value` used
920
+ to put its display string straight into the address — `?tab=License%20Verification%20Status`. It is now
921
+ kebab-cased and lowercased with diacritics dropped: `?tab=license-verification-status`. **If you
922
+ adopted `urlParam` in 0.9.0 with tabs that declare no `value`, your existing links change** — an old
923
+ one no longer matches and lands on the first tab, which is what a stale bookmark has always done.
924
+ Declare `value` on each tab to pin the names for good: a slug is still derived from copy, so renaming a
925
+ tab still moves the link. What `tabClosed` emits is unchanged — still the raw label.
926
+
927
+ - **`ds-auth-layout`'s projected `[brand]` slot no longer forces `width: auto`.** It shipped in 0.9.0
928
+ alongside the slot's `max-height`, but at `(0,2,0)` it out-ranks a width the consumer set on their
929
+ own logo, and with the default `max-height: none` it never did anything: an `<img>` with no CSS
930
+ width already keeps its ratio under `max-height`. If you opted into the max-height AND set an
931
+ explicit width on your logo, add `width: auto` yourself to keep the old scaling.
932
+
933
+ - **The chips of both pickers now render inside a `<span class="chip-row">`** rather than directly
934
+ in the trigger — the element the width measurement needs. **On `[maxVisibleChips]="null"` it is a
935
+ wrapper and nothing more**: `display: contents`, so it adds no box and the chips stay flex items of
936
+ the trigger, identical in spacing, wrapping and height. In the measuring modes — which includes the
937
+ new default — the row becomes a `nowrap` flex box that clips, which is how the cap is enforced.
938
+ Either way, what breaks is a CSS override that walks the structure:
939
+ `.autocomplete-trigger > ds-tag`, `:nth-child` written against the trigger's children, or
940
+ `ds-tag + .autocomplete-input` adjacency — the chips are no longer siblings of the filter input,
941
+ the addon slots or the chevron. Drop the combinator or insert `.chip-row`. Multi-select's trigger is
942
+ `.field-trigger`, not `.trigger`, if you are rewriting a selector against it. Chips carry `data-chip`
943
+ and the marker `data-chip-overflow`; the measurement finds them that way, so leave them in place.
944
+
945
+ - **If you EXTEND a DS component, this is the release that will break you.** `npm run class-surface`
946
+ against 0.9.1 reports **30 exported classes whose member surface grew** — among them `ButtonComponent`,
947
+ `TableComponent`, `ModalComponent`, `SearchComponent`, `AvatarComponent`, `IconComponent`, both
948
+ pickers, every bordered field that gained `controlState`, and every control that gained
949
+ `sizeResolved`. A subclass declaring any of those names now fails with TS4114. **Four classes also
950
+ LOST members**, which is the case a subclass notices differently — an `override` that now overrides
951
+ nothing: `ModalComponent.error` and `.onBackdrop`, `ModalCardComponent.error`,
952
+ `BottomSheetComponent.stop` and `InputComponent.floating`. Subclassing a DS component remains
953
+ unsupported, and this is the third release running where it broke someone, so it is stated in full
954
+ rather than by example. **Two kinds of break the report cannot see**, so check them by hand:
955
+ a member whose SIGNATURE changed keeps its name — `TableComponent.exporting` and `.exportStatus`
956
+ are still there, but as public readonly signals rather than the protected writables a subclass
957
+ could `.set()` — and a plain `private pending: … = null` is not a field the parser recognises, which
958
+ is how `SearchComponent`'s `pending` and `pushed` sit outside its reported seven.
959
+
11
960
  ## 0.9.1 — 2026-08-10
12
961
 
13
962
  ### Added
@@ -85,6 +1034,21 @@ genuine gap worth reporting.
85
1034
 
86
1035
  ### Changed (breaking, behaviour)
87
1036
 
1037
+ - **If you EXTEND a DS component, this release collides with your subclass.** Sixteen exported
1038
+ classes gained members, so a subclass declaring the same name stops the build: `TS4114` wants
1039
+ `override`, `TS2415` fires on a redeclared private, `TS2416` on a differing signature. Not
1040
+ hypothetical — a consumer migrating to 0.8.0 hit exactly this on three `TableComponent` members.
1041
+ **`TableComponent` is by far the largest**: `lazy`, `lazyLoad`, `showLabel`, `showPageSize`,
1042
+ `filterMode`, `columnFilterOptions`, `sortAnnouncement`, `rangeAnnouncement` and `exportFailed`,
1043
+ plus a dozen protected members. **And one signature changed rather than appeared:** `isStriped()`
1044
+ now takes an ABSOLUTE row index instead of a window-relative one, so a subclass that overrode it
1045
+ stripes the wrong rows rather than failing to compile. Elsewhere: both control-value-accessor
1046
+ bases gained `hidden` and `focus()` (inherited by every field control); `ds-input` / `ds-textarea`
1047
+ gained `readonlyBinding`, `isReadOnly`, `minLength`, `pattern`; `ds-modal` and `ds-modal-card`
1048
+ gained `subtitle` and `dividers`; `ds-autocomplete` gained `maxVisibleChips`, `showChevron` and
1049
+ `clear()`. One member was also REMOVED: `TableComponent`'s private `distinctValuesOf()`, now `optionsFor()`, which prefers the new `columnFilterOptions` input. The full per-class list is in `migrations/0.9.0.json`. **Using the `<ds-…>` elements in
1050
+ a template needs no change** — this only reaches you through `extends`.
1051
+
88
1052
  - **`ds-table` `sortMode` and `filterMode` default to `'server'`, were `'client'`.** Server-paged is the
89
1053
  normal case, and there the old default was actively wrong: the table sorted the current page, so the
90
1054
  rows looked sorted and were not — a dev warning existed only to say so. An undeclared table now emits
@@ -97,8 +1061,11 @@ genuine gap worth reporting.
97
1061
  - **The `ds-select` / `ds-multi-select` dropdown panel has a 12px corner, was 8px.** It was the one panel
98
1062
  outside the radius role layer — modal, popover, date picker, time picker, autocomplete, bottom sheet, card,
99
1063
  stat and page footer all resolve `--ds-radius-panel` — so it disagreed with its siblings by 4px and could
100
- not be retuned through the role consumers are told to use. Restore with
101
- `--ds-select-panel-radius: 8px`, or `--ds-radius-panel: 8px` to move every panel together.
1064
+ not be retuned through the role consumers are told to use. Restore with `--ds-radius-panel: 8px`,
1065
+ which moves every panel together. _(Corrected 2026-08-11: this note first offered
1066
+ `--ds-select-panel-radius: 8px` as a way to restore just this panel. That token does not exist —
1067
+ the dropdown resolves `--ds-radius-panel` directly — so setting it does nothing. `ds-autocomplete`
1068
+ does have a per-panel `--ds-autocomplete-panel-radius`; the two dropdowns do not.)_
102
1069
 
103
1070
  - **Seven exit `@keyframes` are gone, replaced by four global presets.** `modal-out`, `ds-sheet-fade-out`
104
1071
  and `drawer-scrim-out` become `ds-fade-out`; `ds-sheet-down` and the four `drawer-panel-close-*` become
@@ -2138,7 +3105,7 @@ rect(…)` → `clip-path: inset(50%)` on visually-hidden helpers, and `min-/max
2138
3105
  set the `--primary-*` roles directly to hand-picked hex.
2139
3106
  - Relicensed from proprietary (UNLICENSED) to MIT.
2140
3107
 
2141
- ## 0.1.0 — 2026-06-12
3108
+ ## 0.1.0 — 2026-06-15
2142
3109
 
2143
3110
  Initial public release.
2144
3111