before / after samplerissueCL 8819998P4 / Blink>LayoutLive max-content width comparison for trailing U+3000 and U+2000, with text-only and inline-span paths plus a break-spaces control.
samplerissueP3 / Blink>GamepadAPICoreHaptics applies HapticIntensityControl as a multiplier on the event intensity, so refreshing an active effect at 0.20 drops output to 0.04. Fix creates the continuous event at intensity 1.0 and drives the level via the dynamic parameter.
samplerissueP3 / Blink>FontsScriptRunIterator anchored script runs on USCRIPT_UNKNOWN (e.g. U+FE1A), dragging adjacent USCRIPT_COMMON emoji and math alphanumerics into a run with no font coverage. Fix remaps USCRIPT_UNKNOWN to USCRIPT_COMMON in IcuScriptData::GetScripts().
samplerCL 8777667issueP4 / Blink>Paintborder-shape only routed solid/double outlines through the shape painter; dotted/dashed/inset/outset/groove/ridge fell back to a rectangular outline. Fix extends PaintOutline with a clipped center-path stroke and inner/outer-half fills.
samplerCL 8746501issueP3 / UI>Browser>WebAppInstallsObserver attached before the alt-error-page navigation commits; aborts/destruction fire DidFinishNavigation with !IsErrorPage(). Fix tracks the navigation id and deletes silently when it never commits; CQ 32/32 green.
samplerCL 8709254issueP3 / Blink>UImacOS PWA Dock badge suppressed by usernoted under notification attribution; fix displays the dock tile and syncs UNUserNotificationCenter badge count.
samplerCL 8549635issueP3 / Blink>MediaPrototype of whatwg/html#12585 plus a11y follow-up: poster img alt exposed as an image node under the video, withdrawn on playback; screencast of the live AX tree.
samplerissueBlink>InputImage with usemap but tap outside every <area> still queued a link-highlight over the whole image; fix skips the image in BestTapNode.
samplerCL 8676164issueP3 / Blink>Image>JXLDecoder output of a jhgm gain-map JXL before and after CL 8676164, with the composited rendition at headroom 1.5/2/4 and a diff.
samplerCL 8666829CL 8766177CL 8766317issueP3 / Blink>Canvas, CaptureFromElementCaptured display-p3 frames were flattened to bt709; the sampler decodes a captured VideoFrame through its own tag and compares it with the canvas color. Stack of three CLs in review.
sampler + firmwareissueP3 / S3 · Blink>HIDFlash the Feather nRF52840 or an ESP32-S3 with the repro descriptor, then check whether the reports keep report ID 1.
indexissueCL 8669858P4 / Blink>FontsUpdateFixed() compared the incoming family against Fixed(), which translates a stored Osaka to Osaka-Mono on macOS, so re-applying the unchanged preference reported a change and invalidated the font cache on every navigation. The list case from the original report was already fixed in b012decc740a9; this is the one instance left
indexissueCL 8663299P4 / Blink>Layout>InlineThe generated break opportunity after a leading preserved space run was emitted at the end of each LayoutText own run, but the run can continue in the next inline box, so the opportunity landed inside it. Not a reuse bug: fresh layout of a pre-wrap block with the run split across a span already collects \u0022 ZWSP ijkl \u0022. The opportunity is now deferred until the sequence ends, flushed in ExitBlock so the offset mapping builder stays consistent, behind the default-on LeadingSpaceBreakOpportunityFix kill switch
indexissueCL 8664319P4 / Blink>EditingCutting everything out of a contenteditable list item copies the ancestor list context, and pasting it back nested a whole ul inside the now empty item, because the destination item is the editable root and the pasted item cannot be placed next to it. ReplaceSelectionCommand now unwraps a pasted single-item list and merges its content into the empty destination, behind the default-on EditingPasteListIntoRootListItem kill switch; a multi-item list and editable-ancestor cases still take the old path
indexissueCL 8656438P3 / Blink>ScrollA snap area that moves under :hover makes the post-layout re-snap pull the scroll offset to its new position, which moves the content out from under the pointer, drops :hover and starts over: the offset oscillates between two values forever. The re-snap still targets the same areas but leaves the offset alone in that case, which css-scroll-snap leaves to the UA, behind the default-on ScrollSnapNoHoverAdjustment kill switch
indexissueSkia CL 1405397Chromium CL 8654518P4 / Skia raster + Ganesh + GraphiteSkia dithers gradients with an 8x8 matrix indexed by device-local coordinates, so a saveLayer dithers in a different phase than a direct draw; background-clip:text paints into such a layer and came out 16120 pixels off by one. Fixed by anchoring the matrix to the global pixel grid. Across the 64 layer origins in an 8x8 window only (0,0) was correct before, all 64 are now, on the raster path, Ganesh GL and Graphite Vulkan: both GPU backends index the same LUT by sk_FragCoord and had the identical defect, each proven red then green against a real driver. Glyphs were out of phase too: Ganesh built the text paint below the device and passed the dither effect a hard-coded zero, so text went 63/64 origins bad to 0/64 once the origin was plumbed through AtlasTextOp. Skia CL 1405397 in review, Chromium CL parked pending the roll
indexissueCL 8653958P3 / Blink>EditingThe bar caret width is chosen in the editable local CSS px and painted through the element transform, so transform: scale(0.1) washes the caret out to nothing and scale(10) paints it ten device pixels wide. The caret rect is now rescaled by the inverse accumulated transform scale, behind the default-on CaretWidthIgnoresTransform kill switch, with a unit test for a transform that changes after the caret was placed
indexissueCL 8650938CL 8654238CL 8654438P3 / Blink>LayoutCanPaintMultipleFragments bailed out on every form control, so a button or select in a repeating table header painted only in the first fragment. Text controls were the reason the check was that broad: they were excluded from fragment traversal, so repeating one trips a duplicated display item id DCHECK. The inner-editor hit test redirection moves into the fragment traversal path, both exceptions go away, behind the default-on RepeatableFormControls kill switch. Sampler has before/after/diff for button, select, input and textarea on screen and in print, plus a second CL giving SVG content one FragmentData per container fragment, which fixes 350858753 (transformed SVG only printed on one page). A third CL lets fragment traversal handle replaced content, which removes the last LayoutObject-traversal fallback and covers 434108536
indexissueCL 8649780P2 / Blink>DevToolsA full-page screenshot makes the browser re-push unchanged WebPreferences, and WebView::ApplyWebPreferences re-applied the embedder font family maps, silently discarding a Page.setFontFamilies override, so screenshot tools got a different monospace font from the second capture on. The font family maps are now skipped when a WebPreferences update does not change them, behind the default-on PreserveDevToolsFontFamilyOverrides kill switch
indexissueCL 8650019P3 / Blink>WorkersWorkerAnimationFrameProvider::BeginFrame wrapped the whole animation frame callback list in one EnqueueMicrotask, so a promise resolved in the first callback only ran after the last callback of the frame instead of before the second one. The callbacks now run directly in the BeginFrame task like on the main thread, behind the default-on WorkerRafMicrotaskCheckpoint kill switch; new WPT also covers the checkpoint after the final callback
indexissueCL 8649429P3 / Blink>PaintTheme parts snapped both edges of their paint rect, so a fractional-size checkbox or radio painted 13 or 14 pixels wide depending on its sub-pixel position and a row of identical checkboxes grew and shrank out of step while resizing. Checkbox and radio now round their size on its own in rendering coordinates; whole-pixel controls and all other parts are unchanged. Default-on ConsistentFormControlPaintSnapping kill switch
indexissueP3 / S3 / Blink>FormsWindows short date "yyyy/MM/dd '('ddd')'" maps ddd to LDML EEE; the date field builders skip day-of-week fields but keep the literal parentheses, leaving "()" in the control. Fix strips day-of-week fields with their surrounding literals before field layout; sampler has live repro plus before/after content_shell captures
indexissueP3 / S3At zoom != 100% the looping time picker columns show a gap where they wrap (00 -> 01): rotation math assumes a 36px cell height but zoomed layout spacing differs; fix measures real cell positions from layout, sampler has live repro plus before/after content_shell captures at 90% zoom
indexissueCL 8644932P3 / Blink>DOMClosing a dialog restores focus to the element focused when it opened; with no such element, focus was left on the node inside the closed dialog where it is no longer focusable, and document key events stopped firing until the page was refocused from the browser UI. The close steps now clear the focused element in that case, which is the else branch the HTML spec leaves out. New WPT red before the fix
indexissueCL 8646510P3 / Blink>SVGSMIL elements were dropped when building a use shadow tree, timed elements inside a use shadow tree were ignored, and the style resolver always took SMIL properties from the referenced element, so animations of an externally referenced SVG never ran. The CL clones the animation elements plus mpath, lets the clones animate their instance, and prefers an instance own SMIL properties, behind the default-on SvgUseInstanceAnimations kill switch. Deterministic WPT reftest that pauses and seeks the timeline instead of sleeping
indexissueCL 8644950P3 / Blink>SVGx/y/dx/dy/rotate lists were consumed per UTF-16 code point, so a combining mark took a list value and was shaped and drawn apart from its base letter. Per the SVG text layout algorithm the lists stay indexed by addressable characters while repositioning applies to typographic characters, so a mark now consumes its value but drops x/y and accumulates dx/dy into the next typographic character, behind the default-on SvgTextPositioningGraphemeClusters kill switch. Two WPT reftests, the second one pinning that later list values do not shift
indexissueCL 8644629P3 / Blink>DOMRange::GetBorderAndTextQuads skipped the children of every fully selected element, including box-less display:contents ones, so their descendant boxes were missing from getBoundingClientRect and getClientRects. A subtree is now skipped only once a selected ancestor actually generates a box, which also avoids emitting a box twice when a display:contents element sits inside a selected element. Existing WPT was red on all desktop platforms, three stale baselines deleted, plus a new internal test pinning both directions
indexissueCL 8630454P4 / Blink>CSSBlink forced table { border-color: gray } in the UA sheet, so a table border and a table-cell border with the same declarations painted different colors; Firefox, Edge and IE11 always used currentColor. The behavior change shipped in M149 behind TableDefaultBorderColorCurrentColor, removable in M153, and trunk is M158, so the CL drops the flag, the @supports not blink-feature block and the two dead branches (+12/-55). Sampler shows 0 px diff on default flags and what disappears under --disable-blink-features
indexissueCL 8626307repro repodup #471204157P3 / Blink>CanvasA canvas from transferControlToOffscreen() posted to a document on another site carries the sending process frame sink ids; the receiver asks viz for that frame sink and the browser kills it with Invalid client ID. The deserializer now drops ids that do not belong to this renderer, so the canvas deserializes as standalone. Browser test across two sites plus a serializer unit test; sampler shows the ping/pong repro on the same build with and without the fix
indexissueCL 8623059demo extensionP4 / Extensions>APImanagement.ExtensionInfo gains a contentScripts field built from the manifest (matches, excludeMatches, js, css), so an extension with the management permission can see what another extension injects instead of parsing locale-dependent permission warnings. Omitted when an extension declares none, and limited to statically declared scripts. Sampler shows the demo extension's panel on the same build with and without the patch
indexissueCL 8625727P3 / Blink>FormsAfter a box selection covers an input[type=range], the press on the slider sits inside the selection, so DragController::DraggableNode reports kDragSourceActionSelection and the armed selection drag swallows the slider drag (not-allowed cursor, value stuck). The fix clears the selection in SliderThumbElement::StartDragging when it overlaps the input, covering both the thumb press and the track press, and leaves unrelated selections alone. Two web tests, each red on its own without the fix. Sampler has live Rendered tabs plus unpatched/patched content_shell captures for thumb, track, no-selection and selection-elsewhere
indexissueCL 6534261P4 / Blink>TextIn Dutch the digraph IJ is capitalized as a unit, so text-transform: capitalize on lang="nl" should give IJsland, not Ijsland. The Dutch rule landed in CL 6534261 but is gated on the ICUCapitalization flag, which went default-on in CL 6600309 and was reverted in CL 6605971 after Speedometer3 regressed 4.34% on win-11-perf (crbug/420998400). Root cause is the title-casing path asking ICU for word boundaries without a reusable iterator, so ICU rebuilds one on every call; Speedometer3's Perf-Dashboard subtest title-cases enough short strings that the setup dominates. CL 8622687 caches the iterator per thread the way the rest of Blink already does. Pinpoint confirms the regression still reproduces on current main without the fix (Perf-Dashboard +3.21% linux, +7.80% mac) and goes away with it (+0.65% / +0.02%), with the overall score flat either way -- so enabling the flag is perf-neutral and the reason to ship is correctness. Sampler shows live Dutch samples next to the expected IJsland output plus the Pinpoint numbers
indexissueCL 8618286P4 / Blink>EditingFrameSelection::SelectionHasFocus() only tied a selection to its text control when the selection was editable, so a readonly input kept painting its selection after the focus moved out of the document (Firefox and IE11 drop it). The change reports no focus for a non-editable selection inside a non-disabled text control, behind the default-on HideUnfocusedReadonlyTextControlSelection kill switch; disabled controls keep the old behavior because they can never hold focus, which FocusingButtonHidesRangeInDisabledTextControl pins. Sampler has live Rendered tabs plus content_shell before/after captures for readonly, editable, disabled and plain-text selections
indexissueCL 8614327Blink>Layout / content-visibilityA content-visibility: hidden holder stops dirty-bit propagation at itself, so when a descendant is dirtied inside the locked subtree the holder gets ChildNeedsFullLayout but ancestors stay clean; on unlock, DisplayLockContext::MarkForLayoutIfNeeded calls SetChildNeedsLayout on the holder which early-returns from the ancestor walk because the holder is already child-dirty, leaving the unlocked subtree unreachable from layout and tripping AssertLaidOut in debug. CL adds an explicit MarkContainerChainForLayout call in that branch behind the default-on MarkAncestorsOnDisplayLockUnlock kill switch, with two unit tests pinning the enabled/disabled behavior. Sampler walks step 0-3 live and ships content_shell before/after captures (identical pixels in release -- the regression is DCHECK-only, pinned by the tests)
indexissueCL 8608947P4 / S5Open since 2019: a column-count:auto multicol with explicit column-width reports max-content = max(column-width, content) instead of column-width per css-sizing-4, so wide children stretch the container; CL clamps max-content to column-width in ColumnLayoutAlgorithm::ComputeMinMaxSizes behind the MulticolColumnWidthMaxContentClamp kill switch and drops the stale multicol-width-005 expectation. CQ on patchset 1 needs 3 legacy baselines bumped (fast/css-intrinsic-dimensions/multicol.html + wpt_internal shrink-to-fit-columns.html) -- sampler lists the exact new values
indexissueCL 8604967P3 / S3MoveToFirstForCulledInline on a descendants cursor scanned every fragment item in the cursor's range per culled inline, so one long line with N culled spans was O(N^2) to hit-test; CL walks LayoutObject descendants directly when the culled inline has <=8 of them (common case), keeps the fragment-scan path for larger inlines (regression guard for crbug/1008523), and adds an IsPartOfCulledInlineBox filter so atomic inlines are skipped. Sampler includes a live elementFromPoint benchmark and before/after content_shell numbers
indexissueCL 8605447P3 / S5backdrop-filter inside nested border-radius + overflow:hidden isn't clipped to the inner ancestor's rounded corners; the hoisted backdrop fill paints into the outermost synthesized render surface where per-quad fast rounded-corner clips can't reach it, so corners show rectangular artifacts; CL folds the inner plain rrect clips into the outermost DstIn mask layer behind BackdropFilterInnerRoundedClips kill switch and re-enables three WPT reftests
indexissueCL 8608187P4 / S5Open since 2016: a near-INT_MAX outline-offset inflates the LayoutUnit rect past LayoutUnit::Max() in the single-rect PaintOutlineRects branch, folding the far edges back into the viewport as spurious red strokes; CL caps the painted offset at LayoutUnit::Max()/4 behind the OutlineOffsetSaturationCap kill switch, sampler shows before/after content_shell captures plus a live iframe repro and a pixel diff
indexissueCL 8599053Hyprland PR 16571omarchy 7559P5 / S3WaylandInputMethodContext::SetCursorLocation sends the caret rect in window-local DIP, but zwp_text_input_v3 expects surface-local logical coordinates (DIP x ui_scale); the sampler marks where the popup belongs vs where the unscaled rect puts it, with a no-CJK validation path via fcitx5 Unicode picker and WAYLAND_DEBUG numbers
indexissueCL 8544570P3 / S3ClearAdvertisementClients() dropped the scanning and watchAdvertisements clients on tab hide without releasing the discovery sessions, so the BlueZ scan stayed active for the lifetime of the tab; sampler starts a scan and logs the hide that should release it, plus a Feather nRF52840 test peripheral that reports the SCAN_REQ rate it sees
indexissueP4 / S4Setting href from the mousedown handler tears down the <use> instance tree, so MouseEventManager drops mousedown_element_ and no click is dispatched; sampler shows a live mousedown/click counter that only stays in sync on a patched build
indexissueCL 8554097P4 / S5Not first-letter related: a space typed under white-space:pre becomes collapsed after the style flips back and InsertTextCommand sweeps it on the next keystroke; live pass/fail checks, the reporter's typing box and a static trailing-space case
indexissueP3 / S3UnitFactorExponentToInt() used a strict > comparison, so the two's complement nibble 0x8 was reported as +8 instead of -8 on every HIDReportItem; sampler includes a live WebHID check plus a Feather nRF52840 Sense repro sketch
indexissueCL 8545991P3 / S3Project page with before/after proof (desktop stored icons, Android install dialog, minted WebAPK), four installable test cases, a desktop reproduction script and a patched arm64 ChromePublic APK
indexissueCL 8519837P2 / S3Sampler for the follow-up: CL 8519837 reported intensity_target as CLLI MaxCLL unconditionally, but 10000 cd/m2 is the PQ nominal peak that encoders write by default and 255 means the field is absent; three identical PQ patches with different stated peaks plus real-world 10000-nit samples show the regression and the fixed fallback to the 1000 cd/m2 default
indexissueP4 / S3LayoutTheme::AdjustStyle() coerced inline and table display values for any element with a non-none appearance value, before the effective appearance was known to be none; keying the coercion off the element's native appearance also fixes the suppressed markers in #408002786 without the table regression of #492945641
indexissueP3 / S3Attribution moved to the pre-paint walk in r1684899, which only visits LayoutObjects, so a display:contents root never had its containertiming attribute read; the sampler compares a contents root against a block root
indexissueCLP3 / S3M154 regression: the popup's aura window copied its hosting widget's DIP/pixel-rounded size, so on a fractional device scale factor the two windows grew a pixel per X round trip and pegged Xorg; dup of #567896504
indexissueP3 / S5OrientationIterator split emoji ZWJ sequences into one shaping run per component because U+200D is not Grapheme_Extend; UAX #29 GB9/GB11 keep the joiner and the joined pictograph in the cluster
indexissueCLP4 / S5Per CSSWG resolution, media skipped by content-visibility should be muted and should not fire media events; sampler checks the scriptable half of that behavior
indexreproissueP4 / S5Clicking into a display:contents contenteditable never focuses it; fixed by treating editable display:contents elements as focusable
live samplerissueP4 / S5 · Blink > PaintWavy text-decoration restarts its wave phase at every inline fragment boundary (nested em/strong); fix anchors the wave at the decorating box. Rendered/Expected/Diff sampler with before/after captures.
live samplerissueP3 / S3 · Blink > InputLong-press a button on a touch screen: :active is retained after pointerup because GestureLongTap never released it; live pass/fail sampler (touch hardware required).
live samplerissueP3 / S5 · Blink > LayoutMeter gauge prints as an empty box unless "Background graphics" is checked; printable live samples plus the UA stylesheet fix.
live samplerissuefirmwarepatched Linux buildP3 / S3 · Linux / Web SerialReproduces the inherited VMIN=0/VTIME=0 DEVICE_LOST failure and verifies the explicit VMIN=1 fix with silent serial probe firmware.
live samplerissueCL 8309055P4 / S3 · macOS / Blink DOMHover targets plus a visibilitychange-driven :hover check to verify hover state is cleared when the tab is hidden via keyboard tab switch.
live samplerissueCL 8227486P3 / S3 · Android installed web appsSwitches between dark and light meta theme colors to verify that the Android status bar follows each runtime update.
live samplerissueCL 8383437P3 / S4 · DevTools NetworkFires requests with distinct Age response headers so custom-column sorting can be verified deterministically.
live samplerissueP4 / S3 · Blink paintReplays the 184px to 185px radius transition and compares the unstable interpolation frame with the corrected corner hull geometry.
live samplerissueCL 8382117P4 / S3 · Blink SVG bindingsChecks known-value normalization and empty-string reflection for missing or invalid referrerpolicy values.
live samplerissueP3 / S3 · Blink Generic Sensor APIShows device-frame and screen-frame gravity vectors in both landscape orientations and flags an inverted screen Y axis.
live samplerissueP3 / S3 · Blink CSS motion pathAnimated dots trace offset-path: border-box on boxes with corner-shape round / bevel / scoop / notch, showing the motion path sticking to the rounded arc instead of the visible contour.
live samplerissueflag-enabled WPT runP5 · Blink image decodersCompares the uncommon grayscale ICC JPEG XL fixture against its PNG reference with live, side-by-side, overlay, and amplified pixel-diff views.
live samplerissueCL 8284219ChromeStatusP3 / S5 · Blink CSSMakes all @property descriptors optional, defaults omitted values to ordinary custom property behavior, and allows typed registrations to use per-use var() fallbacks through a guaranteed-invalid initial value.
live samplerissueP3 / S3 · Blink HTML / CanvasAn install element can render browser-controlled Install or Launch UI, while HTML-in-canvas lets script draw that UI and read its pixels. The fix forces the generic Install appearance whenever the element is in a canvas subtree, masks late installed-state callbacks, and clears existing launch state when the element is moved into a canvas.
indexissuelive reproducerP3 / S5 · Blink CSS / HTML renderingOpen since 2014 with dead attachments and 404 jsfiddles, so the reproducer is rebuilt here. Blink gives <area> display:inline instead of display:none, so an image map area generates a stray empty inline box wherever its <map> sits in the document flow. area.offsetTop then reports that box's position (e.g. 508, the bottom of the preceding div) instead of 0, and offsetParent is non-null. Same root cause as issue 1231263 and the permanently failing WPT html/rendering/non-replaced-elements/hidden-elements.html.
indexissuelive samplerP3 / S5 · Blink Performance APIsAndroid virtual-keyboard viewport changes can recreate the compositor frame source. Presentation feedback from a later, unrelated frame could then resolve an older paint request and inflate INP despite negligible event processing. The sampler repeatedly focuses and blurs native inputs while reporting input, processing, and presentation portions from Event Timing. The fix validates the frame source in PaintTiming and falls back to paint time on a mismatch while preserving legitimate same-source presentation delays.
indexissuelive reproducercustom build (.deb)P3 / S5 · Blink PerformanceAPIsHover events don't stop LCP observation, so e-commerce zoom widgets that insert a larger image on mouseover create a late LCP candidate and inflate the reported LCP. The sampler is a minimal product page with a hover zoom plus a live largest-contentful-paint observer log that flags hover-caused entries. Prototype fix behind the default-off LCPMouseoverHeuristics runtime feature: hover events share one task-attribution context, listener-gated task scopes and :hover style attribution mark hover-caused DOM changes, and paints attributed to hover are excluded from LCP; covered by red/green unit tests in the image/text paint timing detectors.
indexissuelive reproducerP3 / S3 · DevTools Performance panelSearching the flame chart dims every track but only ever matches entries whose text lives in the event title or args. Extension (custom) track entries are synthetic events without args -- their user-visible text sits on devtoolsObj.tooltipText, devtoolsObj.properties and userDetail -- so they were never matched even though they were dimmed like everything else. Fixed in TimelineUIUtils.testContentMatching by adding those fields to the search tokens, with red/green unit tests in TimelineUIUtils.test.ts and TimelineFlameChartView.test.ts.
indexissuelive reproducerP3 / S3 · Blink PaintAn asymmetric concave corner with outline offset or box-shadow spread could calculate a miter beyond its target edge, producing a long stray segment. The fix clamps miter intersections to their corresponding edge segments.
indexissuelive reproducerP4 / S5 · HTML parserAn incomplete markup declaration such as <!DOC> at the end of a closed input stream makes the tokenizer wait forever for the 7-char "DOCTYPE" look-ahead that can never complete, silently dropping the bogus comment and everything after it (empty document / lost trailing text). Fixed in the HTML tokenizer: when a markup-declaration look-ahead returns kNotEnoughCharacters but the input stream is already closed, it now falls through to the bogus-comment path as the spec requires. Gated behind the default-on HTMLParserTruncatedMarkupDeclaration runtime feature so it can be disabled via Finch as a killswitch. Covered by HTMLTokenizerTest red/green unit tests plus a killswitch-off test.
indexissuetest PWAP3 / S5 · Android/WebAPKgetInstalledRelatedApps() returned [] for a PWA on Android 12+. On-device testing (Pixel 6a, Android 16, stable Chrome 150 + local ToT 152): the native "play" path works -- a sideloaded companion app declaring asset_statements for the origin is reported as {id, platform:play, version}. The reporter's WebAPK path can't be exercised here because installing the PWA falls back to a legacy webapp:// shortcut (WebAPK minting fails on this device; local build logs webapk_installer.cc:239 "The WebAPK installation failed"), and getInstalledRelatedApps only matches genuine WebAPK packages. So the symptom is isolated to the webapp/WebAPK path, not the core detection code.
indexissuetest PWA#40911689P3 / S5 · Android/WebAPKsWebAPK install failed when the manifest icon was an SVG with sizes="any". Such an SVG usually has no intrinsic size, so re-downloading it for the WebAPK request with an empty preferred size (WebContents::DownloadImage -> WebImage::DecodeSVG) rasterized to a 0x0 empty bitmap and the request failed. Fixed by deriving a concrete preferred size from each icon's usage before download (Reland "Fix installing WebAPK with SVG icon", crrev.com/c/4544539; desktop dup 40911689 via crrev.com/c/5234019). Added regression coverage: blink WebImageTest.DecodeSVGWithoutIntrinsicSizeUsesDesiredSize and WebApkIconHasherBrowserTest.SvgIconWithSizesAny.
indexissue#510460240test PWAarm64 APKP2 · Android/PWA/Custom Tabs"Open in Chrome" from an installed PWA's in-app browser (Custom Tab) tore down the originating web app: WebAPKs cold-reloaded start_url, shortcut PWAs were removed from recents. Root cause: CustomTabActivityNavigationController.openCurrentUrlInBrowser() finished the activity in the fallback path even for TWA/Webapp/WebAPK. Fix guards that finish() with the existing canFinishActivity check. Verified red/green on a real Pixel 6a (Android 16): unfixed stable Chrome removed the PWA task, fixed build kept it alive.
indexissuetest PWAarm64 APKP3 / S5 · Android/WebAPKInstalled PWAs colored the status bar from the manifest theme_color but left the Android nav bar the platform default. Root cause: ColorProviderImpl.getNavigationBarColor() always returned null AND the nav-bar update only ran in CustomTabActivity (sibling of WebappActivity). Fix hoists updateNavigationBarColor() into BaseCustomTabActivity and reuses the theme color; verified red/green on a real installed PWA (Pixel 6a, Android 16). Default-on killswitch WebAppNavigationBarThemeColor.
indexissueCodePenBlink>CSScorner-shape: bevel; child image covers the beveled border diagonals without overflow: hidden. Analysis with content_shell pixel diffs: border paints identically either way; same overlap class as default round corners -- expected behavior per CSS paint order, recommend Intended Behavior. Author fix: overflow: hidden/clip or matching corner-shape on the image.
indexissuerepro_server.pyP3 / S3Stateful reproducer for a stale start_url on relaunch. Root cause: SimpleCache in-flight entry write lost on Android kill. FIXED on-device on BOTH cache backends (SimpleCache + SQL) with md5-verified on-disk durability: STOPPED-signal flush + open-entry checkpoint. 225/225 deterministic unit tests across both backends. Follow-up: SQL reload-revalidation divergence (separate bug).
indexissueCL 8027830WCO Bug LabP3 / S3Installable PWA sampler: WCO stays enabled in Windows fullscreen, reserving a dead caption hit-test region; fix disables WCO in fullscreen on all platforms.
indexissueP3 / S3Reproducer for fetchLater + pagehide + visibilitychange beacons across the 6 close/shutdown scenarios; closing the last tab of the last window drops the queued request.
indexoriginal attachmentissueCL 7868000P3 / S3pushState abort during an intercepted Navigation API same-document navigation can leave browser progress stuck