The body font-family referenced "Karla", which was never loaded, so
body text silently fell back to system-ui. Point it at the existing
--font-secondary token (Inter + system fallbacks).
Remove the googleapis stylesheet, both google preconnects, and the two
dead fontshare preconnects from the document head. Preload the two
render-critical faces (Inter, Space Grotesk) via Vite ?url imports.
Eliminates two third-party origins and the IP leak to Google.
Replace Google Fonts CDN delivery of the four UI typefaces (Inter,
Space Grotesk, Space Mono, Syne) with latin-subset woff2 vendored into
app/assets/fonts and wired via a hand-authored @font-face stylesheet.
Variable faces keep wght (Inter also opsz). Vite content-hashes the
binaries for immutable caching.
Re-exporting the store singletons (fontCatalogStore, fontLifecycleManager,
FontsByIdsStore) through entities/Font/index.ts meant every consumer of the
barrel eager-instantiated stores and pulled @tanstack/query-core — in dev,
test, and as retained code. Drop the store re-export from the top barrel;
keep the pure surface (types, constants, domain, lib, ui) there for
convenience. Consumers that need stores import $entities/Font/model.
Aligns with the BaseQueryStore carve-out: barrel by default, segment path
when it would drag a heavy or side-effectful dependency.
Breadcrumb is not a business aggregate — it is a scroll-tracking navigation
capability (NavigationWrapper registers page sections into a store), so it
belongs in the features layer, not entities. Move the whole slice and
repoint its three widget consumers. entities/ now holds only Font, a true
aggregate.
BaseQueryStore pulls @tanstack/query-core. Re-exporting it through the
broad $shared/lib and $shared/lib/helpers barrels made every consumer of
those barrels eager-load TanStack at module-eval time (no tree-shaking in
vitest/vite-node), which is what surfaced the queryClient mock init-order
failure. Its single consumer now imports it by path.
The vi.mock factory referenced the top-level QueryClient import. Once
BaseQueryStore landed in the $shared/lib barrel, importing that barrel
for createFilter eagerly pulled @tanstack/query-core into the init
cycle, so the hoisted factory hit the import before initialization
(Cannot access '__vi_import_2__' before initialization). Import
QueryClient inside the factory to keep it independent of module order.
FetchFontsByIds was data-access for the Font aggregate wearing a feature
costume: a single FontsByIdsStore (BaseQueryStore over Font's proxy API),
no UI, consumed only by comparisonStore. Move it beside fontCatalogStore
in the Font entity, collapse its deep $entities/Font imports to relative
paths, expose via the Font public API, and delete the feature slice.
Mock data lived in lib/ and was re-exported through the slice's
production index.ts, advertising fixtures as part of the public API.
Move it to a testing/ segment and drop it from index.ts; consumers now
import explicitly via the $entities/Font/testing subpath.
Repoints the four mock consumers (FontApplicator story, sizeResolver
and fontCatalogStore specs, comparisonStore test) to the new subpath.
Move DualFontLayout and computeLineRenderModel from lib/ to a new
domain/ segment. This is the pure glyph-comparison algorithm — no
framework, no UI, no shared/model dependencies — so it belongs in
domain per the FSD+ ui -> model -> domain interior rule, shielded from
UI-kit and API churn.
Public API is unchanged: the slice index re-exports domain/, so
$entities/Font consumers (ComparisonView SliderArea, Line) are
unaffected.
The space→non-breaking-space swap guarded against a lone space collapsing
to zero width, but the line container now uses white-space: pre (inherited
by Character), so spaces keep their width. The derived had been reduced to
a space→space identity; render `char` directly.
availableWidth was containerWidth minus a flat 48/96px constant, but the
slider track's gutters are responsive CSS padding (px-4/8/12/lg:px-24,
per side). At lg the real padding is 192px while only 96 was subtracted,
so lines were broken ~96px too wide and overflowed the container.
Measure the container's content box (ResizeObserver contentBoxSize) and
use it directly as availableWidth; keep the border box for the slider and
split math. The content box already excludes the gutters, so it tracks the
breakpoints with no constant to maintain.
Extract findSplitIndex; computeLineRenderModel now takes the split index
as a primitive. Line derives its model from `split`, so the $derived
short-circuits on value equality and skips recomputation on spring ticks
that don't move the split (previously every tick rebuilt the model and
re-rendered the line).
Lay the three regions out as inline boxes on a shared baseline. fontA and
fontB now align on the typographic baseline despite differing metrics,
and an always-present overflow:hidden strut pins the line-box baseline so
the line no longer jumps when a bulk run mounts/unmounts or the last
window char morphs to a font of different ascent.
Rewrite Line.svelte to render leftText / windowChars / rightText regions
from a LineRenderModel. Bulk regions render as native shaped text runs so
the browser applies kerning and ligatures; per-char DOM is reserved for
the N-char crossfade window straddling the slider.
Slim Character.svelte: drop the unused proximity prop and the redundant
font-size/font-weight/letter-spacing styles now inherited from the line
container.
Switch SliderArea.svelte to instantiate DualFontLayout and derive each
line's render model via computeLineRenderModel(line, sliderPos,
containerWidth, WINDOW_SIZE).
On initial load, two separate $effects in bindings.svelte.ts — one for
filters, one for sort — each issued its own setOptions with a different
queryKey on the first flush, producing an orphaned
`/fonts?limit=50&offset=0` request immediately followed by the real
`/fonts?limit=50&sort=popularity&offset=0`. Hardcoding the default sort
on the singleton would have papered over the symptom while leaving the
sortStore default and the catalog-store default coupled by hand.
Make bindings the sole emitter of query params:
- features/.../bindings: merge filter + sort effects into one. The effect
reads both stores, builds the merged param object, and issues a single
setParams. No more interleaved setOptions on mount.
- entities/.../fontCatalogStore: gate the observer with `enabled: false`
on construction. The first setParams flips `#enabled` on and triggers
exactly one fetch with the correct queryKey. Removes the need for a
hardcoded default sort on the singleton.
- isEmpty is also gated on `#enabled` so the brief pre-config window
doesn't render "no results" before bindings configures the query.
- The constructor seeds #result from observer.getCurrentResult() because
subscribe may not fire synchronously when the observer is disabled.
The proxy returned `{fonts: null, total: 0}` for empty results, which
fetchProxyFonts surfaced as a generic Error. fontCatalogStore wrapped it
as FontNetworkError, and TanStack retried 3× with exponential backoff —
pinning the loading skeleton for ~7s before settling on an empty list.
Schema mismatches are deterministic; retrying only delays surfacing the
contract violation.
- shared/api/queryClient: introduce NonRetryableError marker class.
The default retry handler short-circuits when it sees this so any
store using the shared client gets fail-fast behavior for free.
- entities/Font/lib/errors: FontResponseError extends NonRetryableError.
- entities/Font/api/proxy/proxyFonts: throw FontResponseError (was a
bare Error). Document that ProxyFontsResponse.fonts is always an array.
- entities/Font/.../fontCatalogStore.fetchPage: preserve a
FontResponseError raised lower in the stack instead of re-wrapping
it as FontNetworkError.
- features/FilterAndSortFonts/api/filters: throw NonRetryableError on
invalid filters payloads and document the array-never-null contract.
Adds an `empty` snippet prop to FontVirtualList and supplies it from the
sidebar FontList. Settled queries with zero results now render a centered
"No typefaces found" label instead of a blank list area.
The seed-defaults effect fired whenever fontA/fontB were still
undefined, including the window between constructor reading storage
and the per-id batch resolving. On a slow batch or fast catalog the
effect clobbered storage with catalog[0]/catalog[N-1], losing the
user's pick on reload.
Now bails when storage already holds IDs, and reads storage via
untrack so per-font selection writes don't re-trigger the effect.
Adds a deterministic regression test that controls catalog/batch
ordering via mockImplementation timing.
The decorative dotted-grid background on the paper surface was a
6-line $derived gridStyle string applied via inline style="" plus four
extra utility classes for color and opacity. Replace with two named
utilities and let CSS handle the responsive switch.
app.css:
- New --color-grid-line CSS var (light + dark) so the grid colour and
intensity auto-switch without consumers needing a dark: variant or an
opacity layer.
- @utility bg-grid (20px cells) and @utility bg-grid-sm (10px cells).
Both reference --color-grid-line, so the same markup paints correctly
in light and dark mode.
SliderArea.svelte:
- Drop the gridStyle $derived block and the inline style= attribute.
- Overlay becomes a single line:
<div class="absolute inset-0 pointer-events-none bg-grid-sm md:bg-grid"
aria-hidden="true" />
Mobile picks the tight 10px grid; the md: breakpoint flips to 20px,
matching the prior JS-driven behaviour with no extra runtime cost.
The slider element had role="slider" and tabindex="0" but no keyboard
handler — the focus ring appeared but the slider could not be moved.
Add a keydown handler implementing the standard ARIA slider contract:
- ArrowLeft / ArrowDown — step left by 1 percent
- ArrowRight / ArrowUp — step right by 1 percent
- Shift + arrow — coarse step (10 percent)
- PageUp / PageDown — coarse step (10 percent)
- Home — jump to 0
- End — jump to 100
Bounds and step sizes extracted as named constants (SLIDER_MIN,
SLIDER_MAX, SLIDER_STEP_FINE, SLIDER_STEP_COARSE). Position updates go
through sliderSpring.target so keyboard moves animate the same way as
pointer drags.
Also adds the missing ARIA attributes that screen readers need:
- aria-valuemin / aria-valuemax (bounds)
- aria-orientation (horizontal)
scale-[0.94] shrinks proportionally — on wide viewports this produced
visibly larger horizontal gaps than vertical ones when the sidebar
opens, and it left the text engine measuring the un-scaled width
(causing the thumb-to-character morph boundary to drift).
Switch to outer-container padding (p-6 when sidebar is open on desktop)
so the paper inherits an equal pixel inset on all four sides. The
ResizeObserver picks up the new dimensions and the layout engine
re-wraps text at the actual rendered width.
Dark mode unchanged. Targets that were reported as "barely visible" in
light theme:
Surfaces / dividers
- --color-border-subtle (light) bumped from rgb(0 0 0 / 0.05) to
--neutral-300 (matches the Input underline variant's border color and
yields a visible hairline on bg-surface / bg-paper).
- New bg-subtle utility (same color as border-subtle but as
background-color) — used by Divider component and the TypographyMenu
inline column separator. Replaces ad-hoc 'bg-black/5 dark:bg-white/10'
and 'bg-black/10 dark:bg-white/10' bands.
- FontSearch + ComparisonView Search wrapper borders switched from
hand-written 'border-swiss-black/5 dark:border-white/10' to
border-subtle so they participate in the palette.
Muted text
- Button tertiary inactive text (light) bumped neutral-400 → neutral-600
(~2.7:1 → ~7.5:1 contrast). Covers the A/B toggle and the font-list
rows in the sidebar.
- Label/TechText muted variant (light) bumped neutral-400 → neutral-600.
Covers the ComboControl value text.
- Link text aligned to neutral-500 / neutral-400 (subtle but visible).
No behavior changes; pure styling.
Ilia Mashkov