microcharts — releases

microcharts 0.11.0

2026-07-07

### Minor Changes - [#88](https://github.com/ganapativs/microcharts/pull/88) [`461a258`](https://github.com/ganapativs/microcharts/commit/461a2585c1e36477514fe9c1679693905d6a6d8b) Thanks [@ganapativs](https://github.com/ganapativs)! - Catalog-wide correctness and accessibility pass over all 106 charts. **Hostile config props no longer reach the accessible name.** A non-finite `width`, `height`, `domain`, `target`, `open`, `window`, `total`, `step` — the kind a host produces with `Number(field.value)` on an empty input — used to render a normal-looking chart while the `aria-label` announced "NaN". Every chart now resolves such a prop to its documented default, and geometry and summary resolve through the same value, so the announced scale is always the scale that was painted. `tests/craft/robust.mjs` grew a config pass that asserts this across the catalog (1255 chart×scenario renders). **A non-finite box is no longer silently destructive.** `viewBox="0 0 NaN 20"` is invalid, so browsers drop the attribute and the chart renders at the wrong scale with a correct-sounding name attached. `chartSide()` in `core/types.ts` clamps it, and both `Chart` (the frame) and chart geometry (the marks) call it, so the two cannot drift. **Theming: a host can now pin light, not just dark.** `[data-mc-theme="dark"]` had no counterpart, so on a dark-OS machine `prefers-color-scheme` won whatever the host said. Hosts that pin light usually override the palette but not `--mc-on-fill`, which left light fills carrying dark knockout ink — measured at 3.0–3.8:1 on real pages, under the 4.5:1 text floor. Added `[data-mc-theme="light"]`, plus a guard asserting every token the dark branch overrides is restated in the light one. **Forced-colors gaps closed.** `--mc-accent` now maps to `Highlight` (roughly a hundred interactive entries draw focus rings as a literal `stroke="var(--mc-accent)"`, which no role mapping reached); the accent element-split is repeated inside the forced-colors block, which lives in a later cascade layer and was overriding it, painting open polylines as solid wedges; `text[data-mc-on-fill]` and `text[data-mc-ink="label"]` take system colors instead of literal rgba. **Summaries.** The paired templates (`Dumbbell`, `Slope`) concatenated an empty percent, announcing "From 0 to 5, up ." whenever the baseline was zero. `round2` returned `±Infinity` for finite inputs past ~1.8e306, so charts announced "∞". `Sparkline`'s interactive entry never passed `strings` to `describeSeries`, so a localized host still got an English accessible name. **WindBarb** could hang: a finite-but-huge `size` made the glyph draw loops iterate ~1e307 times. Bearings are also wrapped into one turn before the trig, which both fixes `direction={5.7e307}` rendering at NaN coordinates and makes 361° draw the same glyph as 1°. Per-subpath gzip grew by a mean of 113 B static / 118 B interactive, entirely from these guards; budgets are re-baselined to match. - [#88](https://github.com/ganapativs/microcharts/pull/88) [`461a258`](https://github.com/ganapativs/microcharts/commit/461a2585c1e36477514fe9c1679693905d6a6d8b) Thanks [@ganapativs](https://github.com/ganapativs)! - `curve="smooth"` no longer paints values the data does not contain. The smooth interpolation was uniform Catmull-Rom, which overshoots: a series dipping back to zero bowed past it, and with `fill` the area crossed the baseline it is anchored to — 1.33 viewBox units on a 20-unit-tall spark, 6.7% of the plot. It now uses monotone cubic (Fritsch–Carlson) tangents, so a smoothed run stays inside the range its own points span. Affects `Sparkline`, `DualSparkline`, `StackedArea`, `RetentionCurve`, `BurnChart`, `HistogramStrip` and `CyclePlot`; those seven subpaths grow by 8–129 B gzip. Also in this pass, per chart: - **Waterfall** — the step connectors are one path instead of one `` per gap: a 100-step waterfall ships 13.2 kB of SSR markup across 107 nodes instead of 26.5 kB across 206. A non-finite `open` no longer reaches the accessible name ("From NaN to NaN"), and the domain scan no longer spreads the level array into `Math.min`/`Math.max`. - **ForecastCone** — the uncertainty bands paint through `--mc-cone-color` / `--mc-cone-opacity` instead of inline `fill`, so they are themable and, in High Contrast Mode, visible at all; they previously carried a `mc-cone-band` class that no rule matched. - **ErrorBudget** — a non-finite `window` no longer reaches the accessible name; the exhaustion cross is one path; the remaining-budget readout carries `data-mc-ink="label"`, so the entrance casts it with the other voice marks. - **Thermometer** — a non-finite `domain` bound falls back to the documented default in both the geometry and the summary, instead of announcing "on a NaN–100 scale" beside a normally drawn tube. - **TallyMarks** — `total={NaN}` drew no marks and printed no overflow numeral while the summary still read the true count; it now falls back to the mark cap. - **ConfusionGrid** — axis initials take a code point, so a label starting with an astral character no longer renders half a surrogate pair. - **EtaBar** — the unrun track uses the same `data-mc-ink="band"` role `Progress` uses, which also gives it a forced-colors mapping. Direct value labels (`text[data-mc-ink="label"]`) now map to `CanvasText` under `forced-colors: active`. They painted a fixed `--mc-neutral` gray against the user's chosen background, which is 3.5:1 on a white Canvas.

microcharts 0.10.0

2026-07-07

### Minor Changes - [#81](https://github.com/ganapativs/microcharts/pull/81) [`e6ccfdb`](https://github.com/ganapativs/microcharts/commit/e6ccfdb4edaef65f329595c94d20090e76d6fa6c) Thanks [@ganapativs](https://github.com/ganapativs)! - The systemic tier of the consistency audit: nine rules the library already stated, none of which anything checked. Every fix ships with the guard that would have caught it, and each guard was confirmed to fail against the old behaviour first. **Annotation labels were laid out at one size and painted at another, on every host.** `annotationFontSize` drives the truncation budget, the `edgeFlip` anchor and every top clamp in `shared/annotations.tsx` — but the `` attribute it wrote is inert, because `:where(.mc-root text)` sets `font-size` as a CSS declaration and a declaration outranks a presentation attribute. The painted size was the HOST's `--mc-label-size`: a different number on each of the 27 hosts, conditional on an unrelated prop on six of them, and on the three that render no text of their own (`PairedBars`, `ControlStrip`, `CyclePlot`) not a viewBox-relative number at all — the `0.75em` default resolves against the surrounding prose, so the same annotation painted at one size in a table cell and another in a heading. The layer now sets `--mc-label-size` on the label itself, so one ramp reaches every host unconditionally. Fixing that made two faults underneath it measurable. Annotation labels clamped their top with a 0.78-ascent model that real faces exceed, so the top label painted ~0.1em above the frame — 1.0 viewBox unit on every host, the same defect and the same fix (`dominant-baseline: central` plus a half-em clamp) that `Sparkline label="minmax"` just landed. And `Marker` was the one label in the file that never went through the truncator: a long label at x = 0 start-anchored and ran the whole width and out of the frame. The containment suite now measures annotation labels on both axes across six hosts, and asserts the other half of the rule — that a label which cannot fit is dropped rather than painted over the edge. **On-fill label ink failed WCAG AA on most of the fills it lands on.** `--mc-on-fill` was a fixed light ink, and one rule spelled its value out instead of reading the token. Measured: 2.1–3.7:1 on the mid-tone categorical fills in either theme, and 2.7–3.0:1 on `--mc-accent`/`--mc-positive` in dark mode, where the palette is deliberately LIFTED and those fills are the lighter of the two. So the ink now flips with the theme, categorical fills take their own `--mc-on-cat` (mid-tone fills want the opposite ink from deep ones), and light-mode sapphire — the one deep cat — is a scoped, documented exception. `TraceFold`'s dim span labels stop borrowing `--mc-surface`, which painted them in the page colour at 1.4:1: not receding, disappearing. `defineTheme` derives `--mc-on-cat` from its own generated palette, so the measurement travels with a custom accent instead of being silently voided by one. **`mono` and `eink` shipped six jewel tones.** Neither preset remapped `--mc-cat-1…6`, so nine categorical charts kept full colour inside a preset whose own comment says "no chroma (the panel can't show it)". Both now derive six tints from `--mc-stroke`. The range is deliberately narrow: a wider ramp puts the middle steps at mid-grey, where NO single label ink clears 4.5:1, and keeping every step light means the ordinary label ink stays legible on all six in both themes. `print` keeps its chroma on purpose — it is a colour-output context — and that decision is now asserted too. **High Contrast Mode could turn missing data into a value.** A hollow mark says so with a literal `fill="none"`, and the forced-colors ink mappings painted straight over it: `HeatStrip`'s muted rect is its MISSING-DATA cell, and it filled solid `GrayText`. One blanket rule now keeps a hollow mark hollow, expressed once so the role added next cannot forget it. Separately, the categorical channel had no forced-colors mapping at all — `.mc-root` sets `forced-color-adjust: none`, which PRESERVES authored hues rather than mapping them, so `--mc-cat-4` shipped near-invisible on a black forced background. Six lightness steps on the system ink now carry the encoding, with a `Canvas` hairline between segments doing the separating that the narrow ramp cannot. **Thirty inline stroke widths opted out of `--mc-density`.** `--mc-sw` is `--mc-stroke-width × --mc-density`, and it was referenced only inside the stylesheet; twenty charts wrote the base token directly and so held their PRIMARY mark at a fixed weight while every stroke around it scaled. Six primary marks were also missing `vector-effect: non-scaling-stroke`, with no CSS default to catch them, so they thickened with their container — and every interactive entry spreads `width: 100%`. Both are now swept and guarded. **A render-phase ref write in the one module all 84 picker charts share.** `useActivePicker` mirrored `selected` into a ref during render, so a render that concurrent React discards (Offscreen, a suspended sibling, StrictMode's double invoke) left the ref holding a value that was never committed and the next tap cleared the wrong selection. The ref is gone: `selected` only changes from discrete events, which React flushes synchronously, so the handler closures are never stale. `react-hooks/rules-of-hooks` is now enforced — verified against a probe file with a conditionally-called `useState` that passed silently until it was pinned — and it immediately caught two missing dependency arrays in this pass. **Inline seats froze at mount.** `useSeatHoist` read the seat once while `Chart` recomputes it every render, and `styles.css` cancels the SVG's own translate as soon as the wrapper is seated — so a post-mount seat change (`SparkBar` flipping bar↔win-loss) left the wrapper at a stale offset with no fallback. It re-reads now, and clears the hoisted copy when a chart stops emitting a seat. **Percent formatting was en-US everywhere.** A literal `` `${Math.round(x * 100)}%` `` is not a percent, it is an en-US percent: most of Europe wants a NBSP before the sign, `tr-TR` puts the sign first, several locales use their own digits. Roughly forty sites across the catalog — painted labels, readout chips, live-region announcements and summary sentences alike — now go through one `makePercentFormatter`, so a `locale` that localised every other number on a chart no longer leaves its percentages in English. `FillWord`, `Hourglass`, `MoonPhase`, `ParetoStrip` and `TimeInRange` gain the `locale` prop they needed to make that reachable. Five more rendered strings that no bundle could translate moved into `SummaryStrings` (`paretoCount`, `biasStripLabel`, `changePointRegime`, `dirNames`, and the quantile band level, which was handed to the bundle as a bare number with `%` baked into the template). **This is the one part of the release that can break you, and only if you supply your own `strings` bundle** — a default `EN` bundle needs no changes. Two things moved: - `SummaryStrings` gained four REQUIRED members. A custom bundle for ParetoStrip, BiasStrip, ChangePoint or Dumbbell will not typecheck until it adds `paretoCount`, `biasStripLabel`, `changePointRegime` or `dirNames` respectively. Copy the EN template from `src/core/strings-*.ts` and translate it. - `bandClause` and `bandEdge` (GradedBand) now receive the interval level as an already-formatted percent STRING rather than a bare number, because the level was the one number on that chart a `locale` could not reach. This one is a runtime change, not just a type error: a template that read `` `${level}% within …` `` now renders `80%%`. Drop the literal `%` from yours. Pre-1.0, so this rides a minor rather than a major — a `1.0.0` here would signal an API stability the library has not claimed yet. **Prose gutters at the digits rate.** `textGutter`'s 0.62 per character is calibrated for the tabular figures the library formats itself; caller-supplied and translated text measures up to 0.95, which is why `textGutterProse` exists — and only two call sites used it. `Hypnogram`, `EventRaster`, `DataDiff`, `SproutRow` and `EventTimeline` reserved room for author text or translated keywords at the figures rate. `DataDiff` also had no horizontal degradation at all, so a long key drove its plot width negative; it gains the drop gate its siblings have. Two charts the audit flagged turned out to be correct — they format their own numbers — and only re-derived the shared estimate inline instead of calling it. Also: `PairedBars` emits `data-mc-origin`, so a negative bar grows out of the baseline instead of inward from the box edge. `StatusDot`'s pulse halo scaled 1.9× from r = 3 in an 8-unit box and painted 1.7 units into the page for the whole animation; the end scale is now pinned by containment, and the looping animation is recorded as the catalog's one documented exemption (the loop is the reading — a monitoring dot holding still says the feed stopped) rather than contradicting the design notes. `MC_EASE_ENTER` had drifted from `--mc-easing`, so a chart's CSS transitions and its scripted entrance eased on different curves; they are pinned equal by a test. (`MC_DUR`/`MC_EASE_MOVE` are kept, not removed — the audit read "no consumers" as dead code, but they are published vocabulary for the UI _around_ a chart, the same category as `PALETTE`/`CATEGORICAL`. They now say so, and say why they are coarser than the engine's per-archetype tables.) **New props, all additive.** Every interactive entry now declares `onActive` as well as `onSelect` — the 19 non-picker scalars implemented only the latter while the shared contract and the quickstart documented both. `ActivityGrid` gains `steps`, a knob its geometry always accepted but its Props never exposed, spelled the way its six siblings spell it. `IconArray` gains `format`, the one chart that took `locale` without it. `RetentionCurve` gains `compare`, with `benchmark` kept as a deprecated alias that still wins when both are passed: `compare` is the catalog word for a second series to read the first against, and `DualSparkline`'s own `compare` JSDoc calls it "the benchmark series", which is how the two names drifted apart. - [#81](https://github.com/ganapativs/microcharts/pull/81) [`e6ccfdb`](https://github.com/ganapativs/microcharts/commit/e6ccfdb4edaef65f329595c94d20090e76d6fa6c) Thanks [@ganapativs](https://github.com/ganapativs)! - Every interactive chart now shows the reading it already announces. Three pickers — `CometTrail`, `IconArray`, `PictogramRow` — lit a hover ring and announced the unit through their live region but painted no readout chip, so a sighted mouse reader got strictly less than a screen-reader user (and less than the documented interaction contract promised). They now paint the chip every other picker does, and `IconArray` finally supplies `datum.formatted`, so its `readout={false}` escape hatch has a string to render. `EtaBar`'s chip was wired to focus only — hovering the bar with a mouse revealed nothing. It now reveals on hover as well as focus, like every other reveal-on-hover scalar. Nine more charts reveal a reading that used to be unreachable with a pointer: `Progress`, `ProgressRing`, `Hourglass`, `BreathingDot`, `HeartbeatBlip`, `FillWord` and `TapeGauge` float their number on hover/focus whenever the mark isn't already printing it (`label="none"`, or a gauge too small for its numeral); `MinimapStrip` floats the window range it was only reporting through `aria-valuetext`; and `TokenConfidence` floats a flagged token's tier and confidence. Empty units read out instead of going silent: `DotPlot`, `MiniBar`, `PairedBars` and `Funnel` show an em dash for a unit with no value — matching `ActivityGrid`, `CalendarStrip`, `RubricStrip` and `StarSpoke` — rather than rendering nothing while the live region announced "no data". Their `datum.formatted` mirrors the chip, per the shared contract. New string tokens (English defaults ship; a custom `strings` bundle can translate them): `iconArrayChip`, `pictogramChip`, `tokenChip`, `heartbeatChip`. `PictogramRow`'s `datum.formatted` now mirrors its chip (`"6 of 8 — 40%"`) rather than the bare percentage. Chip text is the shortest _useful_ reading — never a duplicate of a permanent `label` already on the glyph. Bare counts name their unit (`HeartbeatBlip` → `"3 events"`); `BreathingDot` adds the load band the pulse encodes (`"62% · elevated"`); `EtaBar` floats only what the gutter is not already printing; `Bullet` with `label="both"` floats only the signed gap. `OrbitStatus` / `Thermometer` / `HeatCell` suppress the chip when their permanent label already shows the same number. `readout={false}` suppresses each of these chips and nothing else (playground exposes it for every chip chart), and a new `readout-presence` gate fails the build if an interactive chart ships a picker or a hover reveal without one. ### Patch Changes - [#81](https://github.com/ganapativs/microcharts/pull/81) [`e6ccfdb`](https://github.com/ganapativs/microcharts/commit/e6ccfdb4edaef65f329595c94d20090e76d6fa6c) Thanks [@ganapativs](https://github.com/ganapativs)! - Eight defects found by a catalog-wide consistency audit — each one a rule the library already states and nothing checked. **Labels painted at the wrong size.** `Sparkline label="minmax"` pinned `--mc-label-size` only in `label="last"` mode, so its two extremum labels were laid out at 5–9 viewBox units and painted at the inherited `0.75em` — measured 1 unit outside the viewBox top and bottom, and `.mc-root` is `overflow: visible`, so that is a spill into the page. Both labels now share one baseline (`central`) and one clamp. `TapeGauge` had the same class of bug in reverse: the root pins the _tick_ size, so the hero readout's `fontSize` attribute was inert and the number painted at 7 units while its clearance was reserved for up to 13. **The branded focus ring never applied to anything.** `.mc-root:focus-visible` styles the `