Commit graph

17 commits

Author SHA1 Message Date
Roland
510b24c298 ui: name the directory fields, announce their help text, let long prefixes shrink
Greptile's three findings on #590, all confirmed real before fixing, and each one
verified red against the original code afterwards.

P1, the directory fields had no accessible names.  All five rendered their
visible label as a bare <label> with no htmlFor while passing no label prop, so
their inputs had a description ("/media/wrolpi/") and no name at all -- five
fields in one grid, holding similar relative paths, indistinguishable to a
screen reader.  The labels now belong to the inputs, InfoPopup content and all.
This predates PathInput; the old TextInput call sites had the same gap, and
moving to a component that supports labels is what made it fixable.

P1, description and error were rendered but never referenced.  aria-describedby
pointed only at the prefix, so the help text listing the variables a path
accepts, and the reason a path was rejected, were both invisible to assistive
technology.  Both are now in the description, and an errored field reports
aria-invalid.

P2 was worse than reported: at `flex: none` a long prefix took the entire row
and left the input eighteen pixels wide -- not "consumes the editable area" but
a field nobody can type in.  The media directory is whatever the API reports, so
its length was never ours to assume.  The prefix now shrinks and ellipsises at
half the control, with the full path on `title` and in the accessible
description.

The naming fault could not have been caught where the others were.  jsdom cannot
see layout, and the Settings page needs enough mocking that it had no test at
all -- which is why these five fields were uncovered when they broke.  The new
e2e spec intercepts the API and lets the real page render, so it measures the
names and the geometry a user actually gets.  Reverting one field to a bare
label fails it; a long prefix fails the browser spec at 18px.

962 jest, 15 browser component, 46 e2e, clean cache-busted build, tsc clean.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-07-30 16:10:56 -06:00
Roland
c100ffb378 ui: a browser layer for the library, and the .media reset that never applied
The PathInput overlap could not have been caught by jest -- jsdom computes no
layout -- so the only guard was a grep for one spelling of one mistake.  This
adds the layer that can actually see it, and it found a real bug immediately.

ui-layout.cy.js mounts library components in a real browser through a new
cy.mountUI(), which supplies Mantine's provider and a stamped data-theme and
nothing else.  That spartan setup is the point: these components read no
ambient state, so a spec can assert what jsdom cannot compute -- real geometry,
resolved custom properties, applied filters -- across all four themes.

It reproduces the leftSectionWidth='auto' overlap on purpose and asserts it
still overlaps, because a geometry check that has never gone red is
indistinguishable from one that cannot fail.

The bug it found: the rule resetting `filter` on leaf media inside a `.media`
wrapper scored (0,2,2) against the (0,3,1) rule it was correcting, so it lost
every time and never applied to anything.  Every image inside a filtered
wrapper has been through the luminance matrix twice since that rule landed --
markedly darker than its surroundings, which is exactly what it was written to
prevent.  tokens.test.js asserted only that the rule's text was present, and so
stayed green throughout; it now compares the two selectors' specificity, which
is computable without a browser and goes red on the old selector.

Three guards in ui.test.js keep the library mock-free by construction rather
than by habit: no jest.mock in its own tests, no fetch in any library file, and
no useContext outside the theme controls, whose job is to read the theme.  All
three were verified red against planted violations.

959 jest tests, 14 new browser tests, clean cache-busted build.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-07-30 16:10:56 -06:00
Roland
c534cbfe23 ui: PathInput, and the media directory stops printing over the value
The five Special Directories fields on the Settings page were unreadable: the
media directory printed straight through the value, so Archive Directory showed
`a/media/wrolpi/hain_tag)s/%(domain)s`.

Cause: `leftSectionWidth='auto'`.  Mantine turns that prop into
`--input-padding-inline-start`, and `padding-inline-start: auto` is not valid
CSS -- it resolves to zero, so the input reserved no room for the section
positioned on top of it.  It fails only in CSS, only at runtime, and nothing
warns.

Fixed by expressing the field in the library instead of computing a width.
PathInput puts the prefix and the input side by side in a flex row sharing one
outline, so there is no width to get wrong and nothing to overlap -- the failure
is structurally impossible rather than merely corrected.  It also sizes to any
prefix, which matters because the media directory is `/media/wrolpi` on a Pi and
something else under Docker.

Written test-first, and the tests need no mocks at all -- no context, no fetch,
no router, just the provider test-utils already supplies.  That is the point:
Settings.js needs a pile of provider and API mocks, so its inputs were never
covered, which is how five broken fields shipped.  The tests assert the things
that actually matter and can be checked without layout: the prefix is a sibling
rather than a child, the value never contains it, onChange reports only what was
typed, the prefix is the field's accessible description so a screen-reader user
knows what the path is relative to, and disabled reaches the input rather than
just the wrapper.

Plus a source guard, since jsdom cannot see an overlap: nothing may pass
`leftSectionWidth='auto'` or `rightSectionWidth='auto'`.  It strips comments
first -- PathInput documents the mistake by name, and prose about a mistake is
not the mistake.
2026-07-30 16:10:56 -06:00
Roland
bc5f5d6acb
ui: put the card mimetype accent on the bottom edge
Roland's call, and the better one: along the top the accent sat immediately
against the poster's own top border, so the two edges competed.  On the bottom it
closes the card and reads as a property of the whole thing rather than a header
rule.
2026-07-30 13:53:43 -06:00
Roland
a1fbda7363
ui: bring back the card mimetype accent
Two agents independently dropped it, each because our Card had no `color` prop
and each citing the other's precedent -- which is how a deliberate feature
quietly becomes a convention.  Roland wanted it back.

Card takes a token colour name and draws it as a 3px edge along the top rather
than tinting the surface: surfaces stay flat per the design rules, and an edge
survives night mode without becoming a bright patch.  Because it is a token, the
accent remaps per theme -- four hues in light and dark, a single one in night and
amber, which is the point of those themes.

Restored on every card that shows a file: the video, image and generic file cards
in Files.js, and the archive card.  All of them take it from
`mimetypeColor(mimetype, path)`, the same helper that colours the file icons, so
a card and its icon agree.
2026-07-30 13:51:05 -06:00
Roland
96d84dfab8
ui: the last Semantic UI file, and the stylesheet with it
ControllerPage was the final call site; Theme.tsx then had nothing left to wrap,
so it drops from 861 lines to 338 -- the provider that resolves the theme, stamps
`data-theme` and `data-media-filter`, persists both and hands Mantine our tokens.
Everything below that was Semantic wrappers.

The dark-mode compatibility props went with them.  `i`, `s`, `t` and `inverted`
existed so unmigrated components could borrow Semantic's dark treatment, and
they carried hardcoded greys -- #1B1C1D, #dddddd, #eeeeee -- which is exactly
what the token themes replaced.  Three call sites in Common.js were still
injecting the #eeeeee one into links.

Then `semantic-ui-offline/semantic.min.css` came out of App.js.  That took more
with it than expected, and looking at the running app is the only reason it was
caught:

  Every link lost its colour and gained the browser's underline and visited
  purple.  The dashboard's load averages and section headings were underlined
  and purple.  Links now take `--blue`, which remaps per theme, and underline
  only on hover -- many of them wrap cards, posters and icons.

  `box-sizing: border-box` disappeared, which silently changes the size of every
  padded element in the app.  Semantic's reset had been providing it.

  The navigation bar overflowed a 1000px viewport with `nowrap` and cut "Admin"
  off the right edge -- the one link a user needs to reach the settings that fix
  a broken WROLPi.  Semantic's menu wrapped; ours does now too.

The gzipped CSS bundle is 47.5 kB, and one call site was still leaning on
Semantic's grid classes -- now a flex column, since all three of its callers
pass library components.

The ratchet stays rather than being deleted.  At zero it asserts four things: no
file imports Semantic UI, none calls its toast library, none leans on its
stylesheet, and no file reaches back through Theme.tsx.  It can go when the
packages are uninstalled.

Theme.test.js is deleted, but not before porting the two behaviours it
protected: that Button forwards a ref to the real DOM element (a class-component
ref once crashed the Control page, because portals call `node.contains`), and
that the progress percentage stays readable on every bar colour.
2026-07-30 13:41:05 -06:00
Roland
4b7a195220
ui: three silent library defects the agents' reports surfaced
Each of these was reported by an agent working around it rather than patched in
one file, which is why all three are fixed once, for every call site.

Button accepted an `href` and dropped it.  Mantine renders an anchor only when
it also gets `component='a'`; Semantic spelled that `as='a'`.  So a migrated
call site that kept just the href produced something that looks like a download
link and navigates nowhere -- silently, with no warning.  Button and IconButton
now infer the anchor from the href.

The media filter ran twice on nested media.  A `.media` wrapper filters its
subtree as one unit -- the audio player needs that, because its control glyphs
are SVG and the element-type rule does not reach them -- but the poster image
inside also matched that rule.  Two passes of a luminance-to-red matrix is not
one: the second takes 0.2126 of an already-red pixel, so the poster came out
markedly darker than everything around it.  Leaf media inside a filtered
wrapper is now left alone.

Progress had no indeterminate state.  Semantic's `indicating` animated a stripe
when the size was unknown; without it an upload that has not reported yet sits
at 0% and reads as stalled -- and `aria-valuenow="0"` tells a screen reader the
same wrong thing.  It now animates, reports no value, and falls back to a
static fill under prefers-reduced-motion.

Also: night and amber now override an inline `--label-color`.  A call site may
legitimately need many distinguishable colours -- the Flasher's chip badges use
a twelve-colour palette no four tokens can express -- but on a night-adapted
screen a green or blue chip badge defeats the entire mode.  The labels lose
their distinctness there, which is the right trade.
2026-07-30 13:15:03 -06:00
Roland
87a6a34d68 ui: fix the cypress failures, and two real defects behind them
All 43 cypress specs pass now (was 36 of 43).  Five of the seven failures were
stale selectors, two were real.

Real, and both mine:

  The search modal stopped focusing its input.  Mantine mounts modal content on
  a later tick, so the effect that focused it ran while the ref was still null;
  Semantic's modal mounted synchronously, which is why focusing there had been
  enough.  Fixed declaratively -- the field asks for focus itself, and SearchBox
  also sets `data-autofocus`, which is what Mantine's focus trap honours when it
  takes focus away from a plain autoFocus.

  Escape closed the suggestion list and the enclosing modal at once, so a user
  correcting a mistyped query lost the whole modal.  Escape now dismisses the
  suggestions and stops there; a second Escape, with the list closed, still
  closes the modal.

Stale selectors, all in the specs: Semantic's icon font (`i.search.icon`) is
SVG now, its `.ui.modal`/`.ui.dimmer` are the library's modal and overlay, and
one test asserted a button carried the classes `ui button secondary`.  That last
one is rewritten to check what a user can actually perceive -- that Edit is a
real link pointing at the edit route -- because framework class names say
nothing and break on every library change.

Two accessibility gaps the tests walked into:

  The nav's search control was an anchor with an onClick, no name and no
  keyboard path.  It now has a name, a role and Enter/Space activation, which is
  also what let the spec stop matching on icon-font classes.

  The search dialog had no accessible name at all; it has one now.

One `{force: true}` remains, with its reason recorded: Cypress judges that
field covered because it sits in a fixed modal under a sticky header, and I
verified by hand at Cypress's own viewport that it is the topmost element at
its centre, unobscured, and focused on open.

Also closed a hole in the ratchet that this uncovered.  A file could drop its
`semantic-ui-react` imports, leave the baseline, and still render Semantic
through Theme.tsx's wrappers -- which is exactly how a Semantic modal survived
in the search shortcut long after that file looked migrated.  A test now
forbids it, and the three files doing it (Calculators, KeyboardShortcutsProvider,
and Common.js reaching for theme names) import from the library or from
themes/names instead.
2026-07-30 12:54:24 -06:00
Roland
906705e5b4 ui: thirty-five files off Semantic, leaving fifteen
Phase 2, part 4.  Baseline 50 -> 15.

Migrated: all eight calculators, the rest of inventory and collections, the
Map/Zim/Docs cluster, Videos, Files, Search, Archive, DashboardPage,
DatesSelector, HelpModal, Events, Tags, Flasher's neighbours, admin Settings
and Status, Download, and App.

Also the three shared modules everything imports: api.js and customHooks.js
swap the Semantic toast library for the library's shim (same call shape), and
useForm.js renders the library's Message.  useForm accepts both the old
`{content, header, positive|negative}` shape and the new `{kind, title}`,
because the pages building those objects migrate separately.

Two library gaps closed, both found by call sites needing them:

  ActionInput is a forwardRef now.  Its ref reaches the input rather than the
  wrapper, because callers use it to focus or select the field -- a ref to the
  surrounding div would do neither, which is why one call site had resorted to
  hand-copying the component's markup.

  Header takes a `color` token name.  Three separate call sites had reached for
  an inline style to colour a heading that carries meaning ("3 items to add"
  against "2 to remove"), which is how a hex ends up in markup.

Two files needed repair after being interrupted mid-edit: MapViewer had a
Segment left in its error branch, and VideoPlayer had lost the useContext
import its own code still used.  Neither jest nor tsc noticed; only
`npm run build` did, which is the whole reason that check exists.

Fifteen files remain: Archive, BulkTagModal, Channels, Docs, FileBrowser,
FilePreview, Flasher, Nav, Playlists, Theme.tsx, Upload, VideoPlayer, Zim,
ControllerPage, admin/Downloads.  Theme.tsx is deliberately last -- the
unmigrated files still import its Semantic-backed wrappers, so it can only go
once they have.
2026-07-30 12:54:24 -06:00
Roland
16bdffbac7 ui: fix search suggestion selection, rendering, and clipping
Three findings from review and CI on the Common.js migration, all real.

Selecting a suggestion navigated nowhere.  Semantic handed its callers
`{result}` and every caller destructures that shape, so passing the bare result
left `result.location` undefined -- the search suggestions and the Ctrl+K modal
both stopped navigating.  The shape stays until the call sites can change
together.  (Greptile P1.)

Suggestions lost their custom markup.  Callers pass `resultRenderer` to draw a
tag as a chip rather than a line of text, and the migrated wrapper forwarded it
into a div's props instead of to SearchBox, so every suggestion fell back to the
generic title and description.  SearchBox takes it now.  (Greptile P2.)

The suggestion list clipped its last group.  A cypress spec caught it: with
directories, channels and domains all present, the content runs past the 320px
cap and the final row was cut off.  Eight grouped suggestions is an ordinary
result set, so the cap was simply too low; it is viewport-relative now and
still bounded.

Both regressions are pinned by tests -- one asserting the callback shape a
caller actually destructures, one asserting the custom renderer replaces the
default markup rather than wrapping it.
2026-07-30 11:55:39 -06:00
Roland
851485ef1b ui: Common.js and thirteen more files off Semantic
Phase 2, part 3.  Baseline 63 -> 50.

Common.js is the one that mattered: 51 files import it, so its Semantic usage
was propping up the rest of the migration.  Everything in it now comes from the
library -- Paginator, the two search inputs, DirectorySearch, TabLinks, the
toggles, the messages, APIButton, Breadcrumbs, SimpleAccordion, IframeViewer,
CardPoster, FileIcon, LoadStatistic, the theme menu, and the hotspot icons.

Also migrated: Placeholder, Apps, ArchivePreview, CbzViewer, AddToPlaylist,
TaggedDeleteConfirmModal, the collection edit/tag forms, and four inventory
components.

New in the library, all in the gallery and tested:

  SearchBox     Semantic's `<Search category>` with its grouped results shape
                kept, because four call sites already build it.  A real
                combobox rather than a div that listens for clicks.
  Pagination    Semantic's prop names; onPageChange gets the page number.
  TabBar        the strip only -- routing stays with the caller.
  IconStack     Semantic's IconGroup, named once for assistive technology.
  Table.HeaderCell sorted/onSort -- a real button carrying aria-sort.
  Button/IconButton translate Semantic's size names.

That last one is defensive: unmigrated callers still pass size='tiny', and
Mantine drops a size it does not recognise without a word, so the button
quietly renders at the wrong size.  Mapping here keeps the mixed state honest.

Three real bugs found and fixed on the way, none of them cosmetic:

  * `fileSuffixIconName` returned fourteen icon names the map did not have, so
    opening a folder with a PDF, a Word document or an archive in it logged an
    error and drew a fallback glyph.  All fourteen mapped, with a test that
    walks the mapper's own return values, and another asserting every mapped
    name resolves to a Tabler component that exists.
  * `Statistic` relied on Semantic's `Statistic.Group` for its spacing, so on
    the dashboard the load averages ran together into "1 MIN. LOAD5 MIN.
    LOAD15 MIN. LOAD".  It carries its own spacing now.
  * Semantic's inverted stylesheet colors every descendant of an inverted
    segment, so Toggle labels inherited #dddddd -- white pixels in night mode,
    on the Settings page, which is exactly where someone adjusts their WROLPi
    in the dark.  Restated from tokens.

The baseline ratchet gained a third assertion: a migrated file may still lean
on Semantic's CSS classes without importing the library, which the import check
cannot see and which would collapse silently when Phase 4 deletes the
stylesheet.  There is one such bridge today and it is now declared.
2026-07-30 11:55:39 -06:00
Roland
18dc90b4c2 ui: nine leaf files off Semantic, and the components they needed
Phase 2, part 2.  Baseline 72 -> 63.

Migrated: Tables, SortableTable, DonatePage, Share, Headline, ErrorBoundary,
and the QR code, VIN decoder, and temperature calculators.

New in the library, all in the gallery and tested:

  SearchBox     Semantic's `<Search category>`, keeping its grouped `results`
                shape because four call sites already build it.  A real
                combobox: arrows move, Enter takes the highlighted suggestion
                or submits the raw text, Escape closes without clearing.
  Pagination    Semantic's prop names, but onPageChange gets the page number
                rather than an (event, data) pair.
  TabBar        the strip and its house style only; routing stays at the call
                site, which is what decides what is active.
  IconStack     two glyphs composed into one symbol, named once for assistive
                technology (Semantic's IconGroup).
  Table.HeaderCell gains `sorted` and `onSort`.

That last one is a fix, not just a port.  Semantic's `sorted` prop set
`aria-sort`; migrating a table by hand dropped it, leaving a clickable <th>
that no keyboard could reach and no screen reader could describe.  Putting it
in the library means the next table gets it right without anyone remembering.

Night mode holes closed along the way: the QR codes on the donate, share, and
calculator pages are white by necessity — a scanner needs the contrast — and
they render as SVG, which the media filter does not target.  In night mode that
was a lit panel.  They now carry `.media`, so the filter remaps the block as a
unit and the contrast survives the mapping.

Also: `paypal` joins the icon map rather than being imported ad hoc in a page,
and FileBrowser stops passing Semantic-only table props into a DOM element.
2026-07-30 11:55:39 -06:00
Roland
efe84d17f1 ui: media filtering is a per-theme setting, not a night-mode hardcode
Some checks are pending
Controller Tests / controller-tests (push) Waiting to run
Jest Tests / jest-tests (push) Waiting to run
Filtering media was keyed on `data-theme="night"`, which made it a property of
the theme that no user could turn off.  Some people want their media tinted to
match the interface; some want to see the actual image.

A theme now declares a filter and whether it starts on, and the user decides:

  night   night-red    on by default.  An unfiltered thumbnail undoes dark
                       adaptation, which is the whole point of the mode.
  amber   amber-mono   off by default.  Amber is a look, not a night-vision
                       aid, so media keeps its real colors until asked for.
  light   -            no filter offered, so no toggle is shown.
  dark    -

The choice is stored per theme, so wanting amber's tint says nothing about
wanting night's.  The toggle lives in the theme picker, which puts it on the
Settings page and in the gallery at once.

Mechanically the rules now key on a `data-media-filter` attribute rather than
`data-theme`: a rule keyed on the theme could not be turned off.  index.html
stamps that attribute before first paint alongside data-theme -- without it,
night mode paints a screen of unfiltered thumbnails while the bundle arrives
over a hotspot, costing the user the adaptation the theme exists to protect.
Tests cover both the attribute rule and the fact that the inline script's
duplicated ids and defaults still agree with names.ts.

The amber matrix is the same luminance collapse tinted to #ff9500 (red full,
green 0.584, blue zero) with a gentler gamma than night's: amber carries far
more luminance and clips to a flat wash otherwise.

The gallery gains a Media filtering section showing the WROLPi logo filtered,
beside an unfiltered copy that opts out with `night-unfiltered`.  That copy is
behind a click on purpose -- an unfiltered image is a bright patch, and night
mode should not spring one on someone reading in the dark.

ui-design.md gains the media filtering contract it never got in Phase 0,
including the note that a Phase 3 custom theme names an existing filter id and
cannot ship its own matrix, since that would be code rather than data.
2026-07-30 09:15:04 -06:00
Roland
973a61af28 ui: migrate the extension pages and Configs; five components they needed
Phase 2, part 1.  Three files leave Semantic UI (baseline 75 -> 72), and the
components they needed join the library and the gallery:

  Header       Semantic's shape (as/icon/subheader/dividing).  47 files use
               <Header>, and the level now picks the size, so the type scale
               stays a scale instead of each call site choosing a font size.
  Message      gains an icon and a dismiss button.
  ActionInput  input and button sharing one outline, for <Input action={...}/>.
  Loading      centered loader with a caption, replacing the
               <Loader active inline='centered'> + wrapper div in 19 places.
  Modal        becomes compound - Modal.Header/.Content/.Actions, plus
               Semantic's open, closeIcon, and size names - because 34 files
               are written against that shape.

test-utils mounts MantineProvider with ThemeProvider's exact configuration:
migrated components render Mantine internals and every one of their test files
would otherwise need its own provider.

One Semantic bleed found and fixed: semantic.min.css sets
`h1,h2,h3,h4,h5 { font-family: Lato }` and loads after Mantine's element rule,
so headings ignored the theme's body face.  Only visible in amber, where the
body is mono and the headings were not.  A class-level restatement outranks it;
the rule comes out when Semantic does.

Verified: 925 jest tests, clean production build, and a night-mode audit of
every painted pixel in the gallery (red-only, one known false positive).  The
Configs backup and restore flow was exercised against the real API.
2026-07-30 09:15:04 -06:00
Roland
65d78cae32
ui: night mode actually filters media, and fix three picker icons
Addresses the automated review on #583; all three were real.

The night filter only applied to elements tagged `.media`, and nothing in the
app carried that class - so selecting Night filtered the interface but left
every video, thumbnail, PDF and map full-colour, which is precisely the thing
the mode exists to prevent. The Settings copy promised otherwise, so it was
also wrong.

It now targets the element types instead of an opt-in class, because night mode
has to fail safe: a library of thousands of files reaches the screen through
many components, and one that forgot to opt in would undo half an hour of dark
adaptation. `.media` still works for wrappers that must be filtered as a unit,
and `.night-unfiltered` opts out. The filter is restated on `:fullscreen`, which
a filter on an ancestor stops reaching. Verified on the videos page: all 25
media elements filtered.

Three of the five theme choices rendered a fallback glyph and logged errors,
because the icon map was built by grepping literal name='...' strings and the
navigation picker passes its icon through a variable. Adds sun, moon and
lightbulb, plus a test asserting every theme choice resolves to a real icon.

The pre-paint script wrapped the `matchMedia` fallback in the same try as the
storage read, so a device with storage blocked painted light even when it
prefers dark - recreating the flash the script exists to prevent.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-29 20:08:18 -06:00
Roland
ea50ee2f16
ui: theme picker in Settings, and ship the component gallery
The gallery at /theme-sample is now a normal route rather than a development
one, and Settings links to it, so a user can see what a theme actually looks
like before committing to it. Every new component belongs in the gallery: it is
how the design gets reviewed, and how a component that only looks wrong in one
of the four themes gets caught.

Settings gains a Theme section with the new ThemePicker, which explains what
each theme is for - notably that night mode filters videos, documents and maps
to red as well, not just the interface.

Theme names and picker options move to themes/names.ts so the component library
can read them without importing components/Theme.tsx, which still pulls in
Semantic UI. Theme.tsx re-exports them, so existing call sites are unchanged.
The navigation bar's compact picker now reads the same list, so the two cannot
drift apart.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-29 20:08:17 -06:00
Roland
1bf8ea3da8
ui: token-driven component library on Mantine
Adds src/components/ui: buttons, messages, labels, progress, status, tables,
form controls, panels, cards, statistics, modals, tooltips, tabs, and icons.
Everything resolves its colors through theme tokens, so no component branches
on the current theme, and Semantic UI keeps rendering the app meanwhile.

Mantine is bridged rather than restyled: its own CSS variables are pointed at
our tokens. Four details were not obvious, and each caused a visible defect
in night mode before being fixed:

  - Structural colors must go in the resolver's light/dark blocks. Mantine
    emits its values under [data-mantine-color-scheme], which outranks the
    :root block that `variables` becomes, so its grays were winning.
  - The dark and gray palettes are mapped shade by shade to tokens. Mantine
    reads them directly for table borders, zebra stripes and input chrome;
    they can be neither flattened nor left alone.
  - Mantine writes --button-* as inline styles, which no stylesheet can
    outrank, so the dashed danger treatment sets real properties instead.
  - Switch colors are written onto inner elements, so they need scoped CSS
    rather than a variable on the root. Adds a --knob token for control
    handles, because Mantine's default thumb follows theme.white, which is
    near-black in the dark-based themes.

An audit of every painted pixel in the /theme-sample gallery now finds no
non-red channel in night mode.

Icons come from @tabler/icons-react with the 102 Semantic names the app uses
mapped to components; they stroke with currentColor, so they follow the theme
without any rule of their own. Toasts keep the react-semantic-toasts-2
signature so those call sites migrate by changing an import.

TypeScript 5 is required: Mantine's declarations use `const` type parameters,
which 4.9 cannot parse (Mantine 7 is the same, so downgrading is no escape).
That needs two npm overrides to keep clean installs working, both verified
against a Docker build: react-scripts declares peerOptional typescript ^3||^4,
and fork-ts-checker 6.5.2 assigns typescript.performance.mark, which TS 5
made getter-only - crashing the dev server's type checker on every rebuild.

semantic-baseline.test.js ratchets the removal: it fails if a file not already
listed imports Semantic UI. An ESLint rule would have printed 75 warnings per
rebuild, and CI unsets CI= for its build step, so it would guard nothing there.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-29 20:08:17 -06:00