Flint 0.5 by Chenglong-MS · Pull Request #88 · microsoft/flint-chart

Chenglong-MS · 2026-08-05T07:12:16Z

Theme as a formal specification language

fix: use Math.max for pyramid chart right-bar palette fallback

…0260728-095216 docs: quote pip extras install examples

…aps pages - labelOneBody: operate on the layer host after a facet-channel promotion, and stop copying facet/row/column onto the label layer (Vega-Lite refuses split channels inside a layer; the promoted operator already covers every layer). Fixes assemble crash on faceted rose with data labels (nyt, mckinsey). - applyFurniture: decline on a radius-encoded or faceted arc (radialResistsFurniture). A rose sizes its petals from the panel width, which the vconcat furniture wrapper rescopes (faceted: child_width render error) or collapses (plain: shrunk disc). A theta-only pie is untouched. - Add theme-lab-r2 and theme-lab-gaps playground pages (lazy, family-paged). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

…/450 - ground.ts: dataLabels.show:always now yields to the same band+marks fit facts as whenTheyFit (shared metric, higher marks cap 120), so nyt/mckinsey stop smearing values on bar-n100, pie-25, grouped-color-continuous while keeping pie-10/pyramid-12 labelled. Fixes a first attempt that divided band by series and wrongly suppressed mirrored pyramids. - theme-lab-r2-data: design at 300x300 with a 450x450 stretch ceiling. - ThemeLabR2 page: fixed-width theme-lab tiles in a horizontal-scroll row (no more unreadable 7-col shrink), ScaleToFit charts, pill metadata. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

…t labels, gate showPoints on density line-8 under nyt rendered as dotted smears: a per-series strokeDash (dash as a redundant channel, fired because 8 series > 5 inks) collapsed to stipple on a dense daily path, and showPoints:true drew ~100 haloed points per series that ate the line. - ground.ts: within redundantEncoding:whenNeeded, withhold dash/shape when each series is named at its own mark (seriesEnd/inline legend) — the label already carries identity. - templates/line.ts: withhold the per-observation point overlay when per-series point spacing falls below ~8px; nyt's endpoint-emphasis dots then activate correctly. Keeps points on sparse lines (line-ordinal-30). Same 'preference yields to fit' principle as the always-label floor. Full corpus 0 failures; round-1 lab 121/0; 42 theme tests pass. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

Two cases too particular for a general coverage-round fix: - scatter-color-n50: 50 nominal series exceed any palette; colours cycle and a top legend blows the width out. A real fix (measured legend wrap/fallback + a more-series-than-inks policy) is larger than a patch and risky blind. - pie-25: label smear fixed in iter 2, but palette exhausts past ~24 slices — inherent to a pie past legibility. Gaps page switched to the same fixed-tile horizontal-scroll layout as R2 so the parked cases stay readable. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

…no value and needs no band-end label A violin and a density plot both draw a distribution as an area-mark silhouette. Grounding treated them as ordinary bands: isSummarised (the fact that spares a box plot its data labels) only knew the boxplot mark, so these looked labelable, and bandFamily offered them the seriesEnd placement whose anchor is the band's last reading — meaningful on a streamgraph, an arbitrary near-zero tail on a distribution. Under the direct-labelling houses this stamped a number on the shape (Junior 61660.08, Treatment B 0.00217) and piled the series names at the baseline. Extend isSummarised to key off the chart type (Violin Plot, Density Plot), not just the mark; that one fact drives both gates — the data-label value is withheld and the in-band placement falls through to the house legend. Streamgraphs are not summarised and keep their inline end labels. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

…olumn stops leaking into it The connector span binds a synthetic `__wf_lead` field on x2. With no explicit title, Vega-Lite folds that helper's name into the axis title, which rendered as "Month, __wf_lead". Pin the x-axis title to the real field so the internal column never surfaces in the chart. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

… measure house does not flip the bins A histogram's binned x is an index axis — the reader keys counts off its intervals — but the template never declared it banded, so grounding classed the quantitative field as a measure. A house that seats its measure axis opposite (economist: y on the right) then flipped the bins to the *top* of the plot, where the tick labels collided with the title and no longer lined up with the bars. Add declareLayoutMode marking x banded, matching the bar family and grounding's own stated intent. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

…ayer applyZeroRule converted a unit spec into a layer by hand to append the zero rule, moving the encoding — facet channel and all — into the inner layer. Vega-Lite drops a facet/row/column split left inside a layer's encoding, so any faceted chart whose measure crossed zero (a small- multiple range-area over cities with sub-zero lows) silently lost its faceting and drew every panel's band in one plot as a sawtooth. Route the conversion through appendLayer, which already promotes the split to a real facet operator wrapping the layer. Non-faceted and already-layered charts are unchanged; the zero rule now draws per panel. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

…s out fill marks on white) A sequential ramp whose low end approaches the plot surface makes filled bars invisible — right for a heatmap (cell gaps outline it), wrong for bars. The generalizable principle (a fill mark needs a contrast floor against its surface) touches every sequential-coloured chart and is too risky to land blind mid-round, so it is documented as a reference gap. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

A budget-variance table runs +120 to -80. McKinsey declared only a sequential ramp, so ground.ts painted -80 the palest blue and +120 the darkest -- the reader sees 'less/more', never 'minus/plus'. Every other house declares a diverging ramp; McKinsey did not. A single-hue house cannot diverge, so this is the one place it reaches for a second hue: accent blue below zero, a restrained warm above, its light surface tint at the break. selection.signed set to diverging to match. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

6a: a faceted *table* (facet operator whose panel is an hconcat) rendered blank when a house hung furniture. applyFurniture wraps the chart in an outer vconcat, turning facet>concat into concat>facet>concat; Vega-Lite rescopes the panel width signals out of reach and every column collapses. This is the failure radialResistsFurniture already documents for the faceted disc -- the guard was just too narrow. Added facetedTableResistsFurniture: a facet-of-concat refuses the wrapper and keeps the house rule off, as the radial facet already does. Plain faceted units (line-facet, area-facet) are untouched. 6c: on a line with two encodings (colour=series, strokeDash=actual/forecast), the direct-label houses name the series at each line end and hide the colour key, but the strokeDash key stayed top-right where the end labels now sit, piling into a smear. Added relocateSecondaryLegends: once series are named at the line end, any surviving non-colour key steps to the foot of the plot. Houses that keep a legend are untouched. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

The Waterfall template rebuilds its own encoding objects, and it read only `.field` off `ctx.resolvedEncodings` — discarding the titles the assembler had already resolved from `field_display_names`. The bar layer hardcoded the raw field name, and the shared x encoding carried no title at all. Vega-Lite derives an axis title from the field name of every untitled encoding on a shared scale and concatenates them, so the untitled x also picked up the internal connector column: the axis rendered as "week, __wf_lead". Carrying the resolved titles across fixes both symptoms at once. An explicit title on the primary channel settles the whole shared scale, so the internal columns need no treatment of their own — `title: null` on them would blank the axis instead, since an explicit null anywhere in a shared-scale title set wins. Verified by compiling with vl.compile and reading the Vega axis titles, not just the assembled spec. Swept every bundled gallery case: Waterfall was the only template leaking an internal column into an axis title. 652 tests passing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

`tools/run_full_eval.py` exits with "manifest.json not found" and evaluates nothing. `FIXTURES` still resolves to `packages/flint-py/tests/fixtures/`, but the corpus moved to `shared/test-data/` in 1d1ee39 and this constant did not follow. The report path is derived from `FIXTURES.parent`, so repointing the corpus alone would have carried the report out of `tests/` along with it. Giving the report its own `REPORT_DIR` keeps both outputs on the locations where the committed copies already live: `results.json` next to the corpus, `FULL_GALLERY_REPORT.md` under `tests/`. With this applied the tool completes and reports PASS 476 / MISMATCH 180 / PY_ERROR 0 / JS_ERROR 1 / FIXTURE_MISSING 0. The mismatch count agrees with `pytest -q`. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Two label-overflow fixes found by inspecting rendered R2 audit sheets. bar-n30 / bar-n5 — a house that prints values above its bars clipped the labels on the tallest bars (TV 825, Speaker 749, HDD 707): the bar that defines the scale maximum reaches the plot ceiling and leaves no strip of plot above it for its label. Outer top padding cannot clear it because the title already sits there. Give the measure scale headroom instead, so the tallest bar stops short of the ceiling and every label lands in whitespace inside the plot — the way a labelled bar chart is actually drawn. With headroom the scale-end inside-flip is redundant, so it is dropped on vertical bars (kept on horizontal, where the margin, not the domain, is the remedy). scatter-color-n50 — nyt and economist lay a colour key horizontally at the top, and fifty names in one row run off the side of the canvas, crushing the plot to a sliver. A row has a width, so the entries are wrapped into as many columns as the block holds and capped to a few rows — the same overflow the baseline already resolves, restored to the themed path. All 85 R2 sheets render (0 failures); theme-lab unchanged bar the known exam-ecdf headline note; 694 tests pass; tsc clean. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

…left legends Bar value labels: measure the value-label width against the per-bar slot (band step, or band/seriesCount for grouped/dodged) and, when the label cannot fit, flip a single-series bar's label outside the mark and suppress it for grouped bars (narrow sub-bands with no borrowable padding). Keeps the legend/axis intact. readableOn: pick whichever candidate ink sits furthest from the background in luminance instead of assuming which one is light. Fixes invisible contrast labels on dark houses (powerbi) while leaving light houses unchanged; contrast is a legibility floor, not a fact about which ink a house names "primary". Pie: hide value labels on slices whose arc at the label radius is thinner than one line, via conditional opacity (not a row filter, which would recompute the shared stacked theta and drift the remaining labels off their arcs). Legend: lay a horizontal top/bottom key flush-left with the start-anchored title (anchor:'start', bounds:'full') so the caption begins where the graphic begins, rather than indented to the plot rectangle. Merges with the multi-key row-stacking layout. Updated theme-legend-rows tests to assert the vertical-stacking decision rather than the whole layout object. ThemeLabR2: wrap cells instead of horizontal scroll per section. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

… slopes Series-end labels (seriesEnd/inline placement) ranked each series' rows by the domain field sorted descending, which for an ordinal axis with an explicit sort (months) picks the alphabetically-last value ("Sep"), not the last band drawn ("Dec"). Labels landed mid-plot under the crossing lines. Where the domain carries an explicit sort array, rank by each row's index in it; where the sort is 'descending', rank the field ascending; otherwise keep the descending-field default. Labels now sit in the reserved right margin like a line chart's. Layout: a bump chart has two discrete axes, so the square-cells rule (meant for heatmap tones) pulled both steps equal, giving a tall square panel where every rank change is a near-vertical cliff and twelve crossing series are unreadable. A connected mark is a line, not a grid of cells. For a connected mark with both axes discrete (bars have a continuous value axis; heatmaps declare no cross-section, so neither is caught) skip squaring and instead hold the thin cross axis (rank) to its declared cross-section while stretching the run axis (time) to a bounded 2:1 multiple, clamped to the width budget. bump-many goes 336x336 -> 384x240; the horizontal room untangles the crossings. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

A slopegraph reads in one of two ways, and which one is a house matter. Editorial houses take the value axis away, so the values must be printed on the marks — a "Name Value" label outside each endpoint, no colour key. A house that keeps its measure axis (nature: a journal spine + always-on legend) reads values off the axis and tells the lines apart with the legend, like ordinary Vega-Lite. Printing names on its endpoints only fights the axis for the same margin. - resolveChartDefaults: enable slope end labels (showText/showSeriesInLabel) as a themed default only where the house omits the measure-axis line; a house that keeps its axis is left with its legend. - assemble.ts: create chartProperties for any theme (not only those declaring chartDefaults), or the default never runs. - slope template: drop the now-redundant colour legend once names are on the marks; GroundingContext.namesOnMarks generalises the colour→single-ink drop to any house naming its series on the mark (was hard-keyed to economist). - minBandStep: a chart-specific floor on the per-category band step a house's layout.bandStep may grow but not undercut. slope declares 120, so a compact house (nature's 46px) can't pack the two columns near-vertical with no room for labels. Gates: flint-js tsc clean; 694 vitest; theme-audit 62 sheets 0 FAILED; R2 85 sheets 0 FAILED. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2

…ading Two faults in the name-in-the-band labels a house knocks out of a stacked area at its last reading. 1. The band is an `area`, which Vega-Lite stacks on its own; the label is a `text`, which it does not. Left implicit, every name landed at its series' *raw* value and the whole set piled up at the foot of the plot — the stack might as well not have been there. Naming the band's own offset on the label layer lifts each name to the middle of the band it belongs to. 2. "The end" was found by sorting the domain descending. On a bare nominal index ("Stage 1"…"Stage 12") that sorts as strings, so every series ended at "Stage 9" — the label sat mid-plot AND printed Stage 9's number, not Stage 12's. Rank by declared order where the house gives one, by the field where it is a time or number line, and by arrival order for a plain nominal domain (matching applySeriesEndLabels). The name now sits at the true last column, whose point-scale position is the plot's right edge, so an outside name clears the plot without a fragile width expression that a concatenated view (economist's rule header) resolves against the wrong signal. Gates: flint-js tsc clean; 694 vitest; theme-audit 62 sheets 0 FAILED; R2 85 sheets 0 FAILED. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 33e4b998-4b58-4d7e-9cf8-c107b10e45c2