Repository navigation
Conversation
…face of a font collection.
… decode corruption.
…hey should sit under.
…ts nothing of its own.
…t zipped HTML exports with their web fonts.
…es and accept dropped folders for relative assets.
… renderer painting.
…rs on the path. SVG allows text-anchor on the <textPath> element, where it is inherited by the referenced run, but the importer only read it from <text>. The anchor written on <textPath> was therefore dropped: the run started at startOffset instead of being centred on it, shifting the whole line by about half its length. The textPath's anchor now takes precedence and maps onto Text::textAnchor, matching the browser.
…amily. An inline <svg> is serialized standalone and re-imported without the HTML cascade, so a <text> that omitted font-family inherited nothing and fell back to the importer's default face (Arial) instead of the document's actual font. The snapshot now stamps font-family on the wrapper box, and convertInlineSvg rewrites the serialized root's font-family to the resolved single family (a lone authored family already equals the resolved one, so only CSS stacks are collapsed). A stack left verbatim would be read as one literal family name by the SVG importer, which has no stack resolution.
… are not over-wide. When a viewBox maps onto a different target box, the layer carrying the content transform had no explicit size and measured to its children's unscaled viewBox extent. A parent that sizes to content then laid the layer out by that larger extent — an inline SVG at width:100% inside a flex wrapper came out a scale factor too wide and was clipped to 1200px instead of 1280px. The transform layer (both the merged single-layer case and the wrapper layer) now carries the target width/height.
…s survive import. PAGX can only express a <pattern> fill as an ImagePattern (a tiled bitmap), so a pattern whose content is vector geometry — a stroked wave, a dotted hatch — has no image to tile and the importer drops it, leaving the painted shape flat. The browser can rasterize, so a new pre-snapshot pass draws each vector pattern's tile to a canvas and replaces its content with a tile <image>; the importer's existing image-pattern path then tiles it. Only userSpaceOnUse tiles with an explicit pixel size are handled (an objectBoundingBox tile's size depends on the referencing shape, which the pass does not have); those are left as before.
… it. marker-start / marker-mid / marker-end are inherited presentation attributes, but the importer read them only from the shape element itself, so a marker declared on a wrapping <g> — the common way to mark a whole callout arrow — was dropped and the shape lost its head. The three references now flow through InheritedStyle, and marker content clears them so a marker that inherited its own reference cannot re-expand recursively.
…hem. Browsers paint the check tick, radio dot, range thumb/fill, progress/meter bar, colour swatch and date text+calendar glyph through the UA shadow DOM, which the light-DOM walk cannot reach. `classify` therefore dropped the element entirely (checkbox / radio / range — no text, no authored box visuals) or kept a bare box (color / date / progress / meter), so an imported form lost its checkboxes, sliders and bars. Add a render branch that emits each control's static appearance as inline <svg> from its measured rect and computed styles (accent-color, value/min/max, low/high/optimum, colour value, ISO date). Interactivity is intentionally not modelled — PAGX is a static format — but the visual is the design content, and the missing controls were the largest diff bands on the forms fixture. Also mask password values as one bullet per character rather than emitting the plain value (which leaked the password into the snapshot).
…line. renderTextInput flattened every input/textarea/select value into a single-line span. That is right for <input>/<select>, but a <textarea> soft-wraps its value to the box width (and keeps embedded newlines): forcing nowrap — and, more importantly, emitting neither the box width nor anything that gives the importer a wrap boundary — left it with a bare single-line <Text>, so the value ran off the box as one long line (the "项目描述" value overflowed instead of wrapping to two lines). Pin the span to the textarea's content-box width, keep `pre-wrap` (only when the cascade did not already set white-space), and add `overflow: hidden` — a textarea clips its content, and it is what makes the importer synthesise a wrapping TextBox anchored to the width. The content is top-aligned, as a textarea's is.
… the snapshot.
Both pseudos are painted by the browser but have no DOM node, so the walker never
saw them:
- A drop cap (`.lead::first-letter { float: left; font-size: 96px }`) was simply
missing. Worse, its tall float box overlaps the paragraph's first three line
rects, and the line splitter groups rects by vertical overlap — so the cap and
those lines folded into one run-together span that overflowed the column. The
pass replaces the pseudo with a real floated <span> carrying the cap character
and the pseudo's own box/text styles, drops the character from the text node
(the paragraph then reports clean per-line rects), and removes the
`::first-letter` rules from the page's stylesheets so the pseudo cannot style
the stand-in a second time.
- `<ol>` / `<ul>` markers were dropped outright. Chromium does not expose the
marker text (`content` computes to `normal`), so the pass derives it from
`list-style-type` plus the item's index (honouring `<ol start>` and per-item
`value`) and re-emits it as an absolutely-positioned box pinned to the item's
start edge (`right: 100%`), which is where an outside marker is drawn. The
marker box is sized from the `::marker` computed width/height.
…ur with their ZWJ sequences.
…into the subset. A clamped box hides its overflow behind N line-heights while the DOM keeps every character, so the walker materialised the overflowing lines too — PAGX painted them under the cut with no ellipsis. The line emitter now drops everything past the authored count and rebuilds the last visible line with the browser's "…" ending: the longest prefix that still leaves room for the ellipsis, measured on a scratch canvas so the live DOM stays untouched for the nodes still to come. A `data:` font issues no network request, so nothing carried it into the subset: the computed `font-family` survived on each text node, but the importer had no file to turn into a `<Font file>` resource, and icon fonts (Material Icons ligatures, …) rendered their ligature names as text. Inline faces are now saved next to the output by default — no --download-fonts needed — and re-declared in the subset as local `@font-face` rules. A face whose authored CSS name differs from the file's own family (Playfair → "Playfair Display") stays on the manifest's `--font-alias` route instead, because a `<Font file>` resource registers under the internal name and would never resolve through the rule. Also fixes the family-level de-duplication that dropped the second cut of a family (roman vs italic) when re-declaring faces, and carries the saved faces (style/weight included) through `resolveFontFamilies`/`SavedFont`. Rides along: the `applyLayerName` → `applyLayerAttributes` rename and its test.
Multi-layer url() backgrounds: split two stacked url layers into a synthesized full-bleed child (border-radius inherited, host keeps the bottom layer) and resolve ExtractCssUrl to the first url() function instead of splicing every layer up to the last ')' into one unresolvable path. Table rowspan overflow: hoist painted tr backgrounds into standalone boxes ahead of every row with z-index:-1, replaying the browser's 'row backgrounds paint first' order so a rowspan cell overflowing its row is no longer covered by the following rows' backgrounds. <select>: synthesize the UA chevron (two-tone sizing measured off the reference render) and land the rounded overflow mask on the padding box -- it used to shift by the authored padding and clipped the trailing indicator to a sliver. Date picker glyph: two-tone 12px calendar icon matching the reference render (dark-grey frame, solid black title bar). Tests: new hoist-table-row-backgrounds.test.js and selectArrowMarkup coverage in native-form-controls.test.js; full suite 824 green.
- `form` leaves DROP_TAG_NAMES: a form is a transparent layout container and the walker already renders input/textarea/select/button as plain boxes or text leaves, so dropping the host silently erased every control it wrapped (the Flect landing page's email capture groups vanished from the subset). - The body-root loop in snapshotMainImpl now replays the same paint order renderChildrenInto uses for element children (new shared helper resolveChildPaintEntry), so a `position: fixed; z-index: 100` header authored before the page content no longer sorts by DOM order alone and buries itself under the following opaque sections. - New AUTO_SCROLL_RESCUE_SCRIPT force-reveals boxes that are fully transparent, carry an opacity transition (a reveal wired up but never fired) and hold real content; display:none / visibility:hidden stay respected. The finalize pass additionally waits (bounded by animWaitMs) for the finite transitions the warm-up sweep triggered to settle, so delayed entrance paragraphs are sampled at rest instead of pruned at opacity 0 or frozen as blur ghosts. Verified end-to-end on www.getflect.app (forms, navigation and the HLI copy all survive into the rendered PAGX); npx jest: 55 suites green.
YyZz-wy
requested review from
Hparty,
domchen,
kevingpqi123 and
shlzxjp
as code owners
October 16, 2026 07:44
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## main #3770 +/- ##
==========================================
- Coverage 83.51% 83.47% -0.04%
==========================================
Files 731 737 +6
Lines 105532 107571 +2039
Branches 29218 29989 +771
==========================================
+ Hits 88138 89799 +1661
- Misses 10803 10977 +174
- Partials 6591 6795 +204 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
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
2 participants
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
HTML imports can lose form contents, pseudo-elements, masks and SVG details, paint overlapping layers in the wrong order, or substitute web fonts. This change preserves the browser's measured content and authored font families through HTML and SVG import, PAGX layout, embedding and export.
Validation
mainat7cb58c1db; all 42 rebased patches were unchanged and there were no conflicts../codeformat.sh: passed with clang-format 14;git diff --check: passed.cmake -G Ninja -DPAG_BUILD_TESTS=ON -DCMAKE_BUILD_TYPE=Debug -B cmake-build-debug: configured with TGFX35be4ca.cmake --build cmake-build-debug --target PAGFullTest: passed.cd tools/html-snapshot && npm run build: passed.cd tools/html-snapshot && npm test -- --runInBand: 50 suites and 827 tests passed, including a final run after adding the required copyright notices.Known validation failures
PAGFrameCompareTest.RenderFiles:error_0.pag, frame 310, differs from the existing baseline by one of 73,080 pixels, with a maximum channel difference of 1/255. The baseline and local cache versions for this frame match (60a88e548). A targeted repeat reproduced the failure; its root cause has not been established.PAGXHtmlTest.HtmlScreenshotCompare: ChromiumPage.captureScreenshottimed out forblend_modes_showcase.htmlafter three attempts; the other 78 of 79 batch captures succeeded. A standalone capture of the same 470x420 HTML with the same Chromium launch arguments subsequently passed in about three seconds. This does not make the original batch test pass, and baseline comparison for that batch did not complete.The local baseline cache also reports that it is out of date, so some other screenshot comparisons may have been skipped. No baselines were accepted or regenerated, and no test assertions were weakened. Temporary validation scripts, reports and diagnostic images are outside the repository.
Scope and review notes