v2 - #10
Draft
mihar-22 wants to merge 115 commits into
Draft
Conversation
…ier 3 Regenerates the pnpm lockfile (v9), drops unused dev dependencies, adds jsdom for renderer tests, and updates the import-sort plugin config for v4. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Cue text, class names, and annotations from untrusted caption files could inject markup into the overlay via innerHTML. Text and attributes are now escaped, classes are restricted to word characters, unknown or mismatched end tags no longer corrupt nesting, timestamps render as sibling spans, and numeric/extended entities are decoded. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… long hours Also keeps the full header metadata value (fixes X-TIMESTAMP-MAP), tolerates missing whitespace around the timestamp arrow, and falls back to the internal cue class when no native VTTCue exists (jsdom, happy-dom). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…t tags Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…mbedded fonts Reads Script Info so sizes, margins, and outlines scale with the overlay; renders outlines with text-stroke instead of shadow rings; supports inline override tags (formatting, colour, alignment, position, karaoke), layers, legacy SSA alignment, and decodes embedded fonts. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Each format is lazy loaded as its own chunk. Content-type inference covers the new formats and falls back to the URL extension. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…nc, faster collisions Active cues are found by binary search, cue elements are diffed by identity and inserted in order, collision avoidance jumps past obstacles, percentage positions are honoured, the track language sets lang, layers set z-order, and parts expose a part attribute. Adds syncCaptionsRenderer built on requestVideoFrameCallback. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…s, edge-style presets Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…e cue box Non-region cues never had their position/positionAlign applied, so position:10% size:30% rendered at the left edge. Found by the new real-browser layout tests. Also moves the cqh-based font size off the overlay, since an element can not resolve container units against itself. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ioning tests Upgrades to Vite 8, Vitest 4, TypeScript 7 (moduleResolution bundler), and latest dev dependencies. Layout tests exercise stacking, snap-to-lines, percentage lines, position/size/align, vertical text, styled positions, region roll-up, resize, font sizing, and balanced wrapping in Chromium. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Bundles with rolldown, emits a single bundled prod.d.ts, and mangles private members across chunks with a shared cache via a renderChunk plugin (rolldown's built-in mangleProps only supports single-chunk builds). Adds typecheck, test:unit, test:browser, and screenshots scripts. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…rder stacking Percentage translations in --cue-transform are folded into the collision box so centred cues collide where they are painted. Cues flagged fixed (SSA \pos) are never moved but others avoid them. Top anchored cues now stack first-to-last so simultaneous cues read top-down in cue order. Renders are no longer dropped while a resize debounce is pending, which previously hid cues added during async parsing. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Cue boxes use max-content width bounded by the play-resolution margins instead of spanning the full margin width, so unrelated cues no longer collide. Centre alignment uses a translate transform, and \pos cues are marked fixed. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The sandbox now hosts named scenarios (cues, regions, region-scroll, collisions, ssa, edge-styles) and scripts/screenshots.mjs renders them with Playwright into assets/. README embeds the new images. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The parser collects STYLE blocks (result.styles, onStyle callback) and recovers when the blank separator is missing. The renderer rewrites ::cue, ::cue(selector), and ::cue-region selectors to the rendered DOM, scopes them to the overlay via data-scope, and restricts declarations to presentational properties with no external resources. Also adds the SCC channel option type used by the upcoming CC2 support. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…rances Signature must be WEBVTT followed by space, tab, or end of line (BOM allowed); a missing signature is reported but tolerated outside strict mode. Strict mode enforces the spec timestamp grammar and requires % on percentages, while lenient mode accepts comma separators, bare numbers, and legacy align:middle. A timing line inside cue text ends the current cue. Voice annotations collapse all whitespace and surrogate character references decode to U+FFFD. Adds WebVTT file and cue-text conformance suites. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…nd inference Reads ttp:cellResolution for c units on origin/extent/fontSize, emits scaled font sizes for %, c, px, and em, maps tts:writingMode tbrl/tblr to vertical cues with swapped line/position axes and displayAlign handling, and infers a paragraph end from the next sibling before falling back to the 10s default. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… percentage, and RTL cues Resolved starting boxes now carry width and height; previously they were NaN, which silently disabled lineAlign for percentage lines and any math using the box size. Percentage lineAlign center/end shift the box back instead of forward, vertical-rl cues count lines from the right edge, negative lines anchor the box's far edge so line -1 sits flush, and auto position for start/end follows the base direction in RTL. Adds a spec-rendering browser suite measuring all of this in Chromium. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… boundaries Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… nested Adds an SSA/ASS conformance suite covering colours, alignment schemes, layers, margins, override tags, karaoke, and embedded fonts. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…and development workflow Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…scenarios Uses Vitest's toMatchScreenshot with pixelmatch and a 0.5% mismatch allowance. Baselines are stored per browser and platform; the committed set is chromium/darwin, and a first run on another platform records its own. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… cc_data parsing Extracts the SCC parser's CEA-608 core into CEA608Decoder, which accepts byte pairs per field or cc_data triplets from A/53 user data and SEI messages, so CC3/CC4 on field 2 are decodable from video streams. parseCCData extracts triplets from GA94 user data. Published as a separate media-captions/cea entry so the core bundle stays small. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Assembles DTVCC packets and service blocks from cc_data, decodes C0/C1/G0/G1/G2/G3 code sets, models the eight caption windows with pen attributes, and emits one VTTCue per visible window with anchors mapped to line/position. Exported from media-captions/cea alongside the 608 decoder. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
`BROWSERS=chromium,firefox,webkit` selects the Playwright instances (default Chromium); CI runs all three. Screenshot baselines stay Chromium-only, so the visual suite skips elsewhere. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…aptions/renderer entry
`CaptionsRenderer` did everything in one class, so an SRT player shipped region handling, SSA
typesetting, and animation code it never used. The renderer is now a small core plus features:
- `CaptionsRendererCore` / `createRenderer(overlay, { features })` owns the track, cue elements,
WebVTT positioning, collision avoidance, timed text, and cue events. Features hook into its
phases (`setup`, `changeTrack`, `createCue`/`disposeCue`, `containerFor`, `beforeMeasure`,
`measureContainer`/`writeContainer`, `writeCue`, `update`, `reset`, `resize`, `destroy`); hooks
run in phase order, never array order, and same-named features replace each other.
- Features: `regions()`, `typesetting()` (layout, text style, layer, clip rects), `animations()`,
`vttStyles()`, `announcer(mode)`. `defaultFeatures()` is the first four.
- `CaptionsRenderer` is the core with `defaultFeatures()` plus the announcer when `announce` is
set; `init.features` overrides the set. Its public API is unchanged and every existing suite
passes against it in Chromium, Firefox, and WebKit.
- New `media-captions/renderer` entry exports the core, features, and `syncCaptionsRenderer`.
Tree-shaken and minified, the core is ~8.5 KB gzipped against ~11 KB for the full renderer.
- Parsers stay decoupled: they only fill the cue model. In development the core warns once per
capability a cue or track uses that no installed feature handles, naming the feature to add.
- The core no longer reads `cue.layout` to decide WebVTT sizing; it checks whether a feature
already sized the box, so an uninstalled typesetting feature really is ignored.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Summing chunk files no longer says anything useful now that the renderer is composable, since a consumer's bundler drops the features it does not import. `pnpm size` bundles small probe modules against the built package with rolldown (parsers external, minified) and budgets those: parse dispatch, renderer core, core + regions, full renderer, a typical player, and everything. Published opt-in entries keep their as-shipped budgets. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A `features` option (URL `?features=core|core-regions|core-typesetting|...`) rebuilds the stage with `createRenderer` and the chosen features, so the effect of each feature and the dev warnings for missing ones can be seen against every sample. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
IMSC image cues set `background-image` on the cue display box, where the stylesheet had no sizing rules, so small images tiled across the region. Found by comparing the DOM output with the canvas writer; the intent (`contain`, no repeat, centred) already existed for `--cue-bg-image`. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…s/canvas) Abstracts the measure phase away from the DOM and adds a second writer that paints into a 2D canvas, reusing the cue model, `CueTrack`, stacking order, and the pure layout engine unchanged. - `TextMeasurer` is the only thing layout needs from a surface (`canvasTextMeasurer`, `monospaceTextMeasurer` for tests). `measureCue` flows cue text (greedy wrap, balance pass, character overflow, inline tags, class colours, span styles, timestamps) and produces the same `CueLayoutInput` the DOM measurer does; the browser suite checks both writers' boxes agree within a couple of pixels. - `paintCaptions(ctx, cues, time)` is stateless (WebCodecs, OffscreenCanvas, thumbnails); `CanvasCaptionsRenderer` wraps it with a track, time, options, and an image cache, and works with `syncCaptionsRenderer`, whose parameter is now anything with `currentTime`. - Renders WebVTT positioning, regions with roll-up, the SSA/TTML/CEA-708 layout and text style model (boxes, anchors, strokes, shadows, rotation/scale about the anchor or `\org`, clip rects and polygons, `\p` drawings via `Path2D`, image cues, outlines), `cue.animations` sampled at media time (numbers, lengths, colours, transform lists, length lists), and karaoke sweeps as a clipped double fill. - `css-values.ts` bridges the model's CSS strings (a small `calc()` evaluator, transforms, shadows, strokes, clip paths) until the style model carries typed values. All regexes are linear (the ReDoS audit covers the new module). - Known gaps: vertical writing (drawn horizontally), ruby positioning, 3D rotations, blur, STYLE blocks. Documented in the README and docs/design. - Size probe: the canvas renderer is ~12 KB min+gz tree-shaken. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
What the canvas writer showed about the parser/renderer contract: the layout engine was already the right boundary, the DOM-specific workarounds (calc() strings, inline-block transforms, data attributes read back during measure, rAF before region activation, write-time clip resolution) disappear on a canvas, and the CSS-string style model forces a second interpreter. Proposes typed lengths, colours, transforms, strokes, shadows, fills, and keyframes with writers as serialisers, and a three-step additive migration. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A `Canvas` tab paints the current sample through `CanvasCaptionsRenderer` into a device-pixel-ratio canvas on an identical stage, driven by the same media time and styling controls, so the DOM and canvas writers can be compared frame for frame. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The cue model carried CSS strings (`fontSize: 'calc(var(--overlay-height) * 0.0667)'`,
`transform: 'scaleX(1.2) rotate(-10deg)'`, `clipPath: 'polygon(calc(...))'`, karaoke gradients,
Web Animations keyframes with CSS property names). That leaked the DOM writer into every parser
and forced the canvas writer to parse our own output back into numbers. Values now live in the
model and each writer serialises them.
Model (`src/vtt/vtt-cue.ts`):
- `CueLength` (`px` | `vw`/`vh` of the overlay | `em` | `%` of the box), `CueTransform` with
`origin` (box %) or `originAt` (overlay point, SSA `\org`), `CueStroke`, `CueShadow`,
`CueSweep` (karaoke fill colours), `CueClip` (`rect`/`polygon` overlay-relative, `inset`
box-relative), typed `CueKeyframe`.
- `CueTextStyle`: `fontWeight` number, `italic`/`underline`/`strike` booleans, `wrap`,
`stroke`/`shadow`/`outline` objects, `padding`, `transform`, `image { url, fit }`.
- `CueSpanStyle`: same value types plus `blur` and `sweep`; the `display: inline-block` hack is
gone (the writer adds it when a transform needs it).
- `CueLayout.clip` replaces `clipPath` and `clipRect`.
Writers:
- `src/vtt/style-css.ts` serialises for the DOM and string renderers: lengths to `calc()` over the
overlay variables, transforms with origins, strokes, shadows, sweeps as clipped gradients,
keyframes to Web Animations properties, and clips either as pixel insets/polygons once the box is
laid out or, for animated clips, relative to the box at each keyframe.
- `src/canvas/values.ts` resolves the same values to pixels; `css-values.ts` (the CSS bridge) is
deleted. Animation sampling interpolates typed values field by field.
Parsers: SSA (`_lenY` returns a `vh` length, `\clip` is overlay percentages for positioned and
layout-positioned cues alike, scroll bands are a fixed rect, sweeps are a 0..1 keyframe, `\org` is
an overlay point), TTML (lengths, strokes, first shadow, opacity, `image`), CEA-708 (`em` strokes
and shadows, inset wipes, numeric marquee), MicroDVD (`em` sizes, `strike`).
Tests updated to the typed values; corpus snapshots regenerated; both writers pass their suites in
Chromium, Firefox, and WebKit. Size: the canvas renderer lost ~1.1 KB, the DOM renderer gained
~1.2 KB (it now serialises), and the SSA/TTML parser chunks shrank.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…d out Re-measuring after a resize (including the initial ResizeObserver callback) read cue boxes at their already laid out positions and applied the line offset again. Horizontal cues were accidentally protected because their written `--cue-top`/`--cue-bottom` counted as an explicit position; vertical cues were not, so every vertical percentage-line cue drifted to the right edge within a frame of appearing. The explicit-position check now covers left/right for vertical cues and the layout engine keeps such boxes on their block axis. Found by comparing the DOM and canvas writers on the TTML vertical sample. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
`vertical: 'rl' | 'lr'` cues flow into columns one line height wide, read right-to-left or left-to-right, with `text-orientation: mixed` semantics: CJK and fullwidth glyphs stand upright one em each, other scripts rotate a quarter turn and advance by their width. WebVTT `position`/`size` run along the height, the padding axes swap like the stylesheet, and the layout input carries the writing direction so the shared engine snaps columns along x exactly as it does for the DOM. The browser suite checks the column box matches the DOM box and that glyphs are stacked. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ted captureStream The Canvas tab gains two buttons. A compositor draws the mock video frame and the captions canvas into one canvas, streams it with `captureStream()` into a hidden `<video>`, and enters picture-in-picture or fullscreen (`webkitEnterFullscreen` on iOS). That is the route by which custom captions reach surfaces that only show video pixels; a player would draw its own `<video>` or a WebCodecs frame instead of the mock. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs/MIGRATION.md lists every breaking change on this branch with its replacement: ESM-only and the new entry points, spec-strict WebVTT parsing, the typed cue model, the composable and canvas renderers, and the token renderer signatures. The README links it, shows the captureStream compositing recipe, and drops vertical writing from the canvas writer's gaps. `lengthToCSS` and the other serialisers are exported from the main entry for consumers who read the typed model. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Every CI run on this branch failed at setup with "No pnpm version is specified". Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The frozen install in CI refused the lockfile because its overrides section was missing. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The 20k-cue stress inputs took ~2.5 s under coverage on GitHub's runners against a 2 s budget. Regressions the budget exists to catch are orders of magnitude, so CI gets 8 s. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
`pnpm docs` is forwarded to npm's built-in `docs` command, which tries to open a browser and fails on CI. The TypeDoc entry points now include the mp4, renderer, and canvas entries. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Ruby bases flow as one unbreakable run carrying their <rt> text at half size; the annotation is painted in a band above the line (beside the column in vertical text) and the line grows by that band like the browser's ruby line box. Transforms are flattened to a 2D affine matrix so rotateX/rotateY project orthographically, as CSS does without perspective, instead of being dropped. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
An end tag only closes the current node (</ruby> also closes an open <rt>); one that matches an ancestor is ignored instead of closing through it, so <b><i>x</b> y keeps y bold italic as browsers render it. Timestamp nodes stay transparent to end tags. Flips the four remaining WPT cue-text divergences. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
lenient: false applies the WebVTT grammar strict mode enforces (no bare percentages, no align:middle, three fraction digits, any --> line ends a cue, a bad signature gives up on the file) while keeping the recovery of the default mode: invalid cues are dropped and reported, never thrown. The WPT harness now parses every file that way, so the six file-parsing divergences that were deliberate lenient-mode tolerances pass and no test.fails remain. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Recorded on ubuntu-latest by record-baselines.yml (run 34150680622); the workflow goes back to manual dispatch. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…uite Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Chromium grows a ruby line box by about a pixel with half-size annotations; the annotation overflows into the cue padding instead. Match that so the canvas and DOM boxes agree on every platform, rather than growing the line. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Chromium overflows the annotation; Firefox and WebKit grow the line box by about the band. The canvas follows Chromium, so assert exact agreement there and bounded growth elsewhere. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The frame loop advanced lastTime before checking whether the canvas needed a repaint, so the canvas only updated when cues changed and stayed frozen at the initial time while seeking or playing. Decide once and reuse for both writers. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
globalAlpha applied per draw call, and a glyph is painted up to four times (background, shadow, stroke, fill), so a cue at alpha a composited to 1 - (1 - a)^n: fades held near opaque and then dropped off, which read as transitions cutting out early. Groups below full opacity now paint at full alpha into a pooled scratch layer that is composited once under the caller's clip. Applies to cue, region row, and per-run opacity. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
A ground-up rework of the captions library, shipped as a new package. Breaking changes are intentional and marked with
!in commit subjects;docs/MIGRATION.mdlists each one with its replacement. Every commit is self-contained and the log reads as the changelog.What is in here
Security and conformance. Cue text is built as DOM nodes rather than HTML strings, closing stored XSS through entity-encoded text, class names, and voice annotations. The web-platform-tests WebVTT suites are vendored and run in CI: all 118 converted tests pass under the new
lenient: falseoption (the spec grammar with browser-style recovery); the default lenient mode keeps its documented real-world tolerances. Every regex literal insrcis audited for super-linear behaviour, nesting is capped, and a seeded fuzz suite runs against every parser.Formats. VTT, SRT, SSA/ASS (full typesetting: override tags,
\pos/\move/\fad/\t, karaoke,\pdrawings,\clip,Effectfields, embedded fonts), TTML/IMSC (with the vendored IMSC test documents), SCC and live CEA-608/708 fromcc_data, LRC, SBV, SAMI, MicroDVD, and fMP4wvtt/stpptracks.Typed cue model. Positioned formats fill
cue.layout,cue.textStyle,cue.spans, andcue.animationswith values (typed lengths, transforms with pivots, strokes, shadows, clips, sweeps, keyframes), never CSS strings. Cues round-trip through JSON. Writers serialise the model: the DOM renderer to CSS, the canvas renderer to pixels.Composable renderer.
CaptionsRendereris a small core plus features (regions,typesetting,animations,vttStyles,announcer) that tree-shake when unused;createRendererfrommedia-captions/rendererstarts from the core. Measure, layout, and write are separate phases with a pure layout engine, transform-aware collision avoidance, a liveCueTrack,enter/exitevents, reduced-motion handling, and a screen-reader announcer.Canvas renderer.
media-captions/canvaspaints the same cue model into a 2D context through a headless measurer that agrees with the DOM within a couple of pixels. Horizontal and vertical writing, regions, typesetting, animations sampled at media time. StatelesspaintCaptionsfits WebCodecs andOffscreenCanvas; the compositing recipe for picture-in-picture and iOS fullscreen is in the README and the playground.Tooling. Vite+ (oxfmt, oxlint, vitest, tsdown) with publint and attw package checks; real-browser layout suites in Chromium, Firefox, and WebKit; screenshot baselines (Linux ones committed, so the visual suite runs in CI); size budgets from tree-shaken usage probes; TypeDoc; a playground covering every parser and renderer feature.
Sizes (minified + gzipped, tree-shaken, parsers external)
CaptionsRendererReview notes
docs/MIGRATION.mdfor the API delta,docs/design/canvas-and-style-model.mdfor why the model is typed.tests/wpt/KNOWN_DIVERGENCES.mdfor the lenient-mode tolerances and the parser bugs the WPT suite found.pnpm dev,/playground/index.html) shows every sample in the DOM and canvas writers side by side.🤖 Generated with Claude Code