Known limitations

Destroying a Window on Linux/macOS does not release GPU resources either: GLX contexts, Metal devices, and GrDirectContext remain until the thread exits

Stabilize Windows LCD paragraph eligibility and capture phase

On Windows, normal-line strut ascent/descent and text baselines inside atomic boxes retain fractions, with whole-line rounding used to match Chrome integer line heights

Whole-word protection splits on whitespace and ASCII punctuation, so segments with other break opportunities, including ideographs, hyphens, slashes, and Thai, are protected as a whole or split incorrectly

Align the generated-content inline start in a leading block-in-inline fragment

When the first word does not fit on the final line with text-overflow:ellipsis and legacy line-clamp, KUI retains the WPT webkit_line_clamp_036/F-WPT-FIX-924 behavior of falling back to a soft-wrap opportunity and leaving only an ellipsis, while Chrome 153 on the same machine truncates by grapheme, as in supercali followed by an ellipsis

Unify line-height:normal union rounding for mixed custom/fallback fonts and split inlines

Resolve fractional-phase and baseline behavior for font-size-adjust with unicode-range and system fallback

Register known differences from Chrome that have not met the issue-registration threshold; 34 old T-BUG entries have been moved out of this table

Width-dependent `break-spaces` corrections relayout the entire paragraph for each forced break; the same path also searches U+2028 correction fragments code point by code point for same-line membership

CPU blend-layer fallback cases (DstIn, masks, filters, shadows, non-translation text, wavy lines, and unknown operations), GPU, and PDF still use full-size large layers

Every box adds save/translate/restore operations to the display list even if it paints nothing

The performance experiment matrix lacks realistic application structures and budgets independent of scale

Consumer projects participate only in SDK release verification and do not collect interaction performance traces

Automatically find cases where one dirty box causes workload to grow with scale by generating combined structures

Opening a small overlay or form for the first time adds 6-9 ms of compositing time

Fixed-position elements bypass damage culling; visible-overflow containers still record their own backgrounds and borders outside the damage region

On the GPU, the first partial replay frame with many opacity layers still costs about 5ms more than a full repaint

Paragraphs beside floats are laid out at full width first, then laid out again at the available gap width once stability is established

Changing the width every frame with legacy soft-line clamping adds line-end probes and paint reconstruction work

Some Lua OffscreenCanvas drawing operations still use the slow general RGBA raster pipeline

Business dialogs in 3.61.3 still visibly stutter on first open and replay

The cause of historically slow width changes in the unbroken plain chain remains unidentified

Paragraphs with atomic inline boxes in documents without a backend waste a full-paragraph attempt before falling back to individual dispatch on every layout pass

Governance of offline performance gates

Performance-module test debt

Drift in performance-module documentation and contracts

Experimental-matrix workload ceilings and local timing baselines

Text containing ASCII spaces still requires a separate paragraph probe for min-content

When dragging the width, a grid inside a flex container receives a full layout, including a minimum-column trial, during flex-basis measurement and is measured twice per frame

Remaining paragraph reflow during steady-state width dragging (center/right alignment anchors and soft wrapping), plus the first-frame memo-table storage gate

Root font-size changes restyle all descendants and fully rebuild layout regardless of whether rem is used

Paragraphs requiring placeholder permutation into visual slots incur another layout pass, repeating the early function work and run population

Remaining full-tree walks during selective descent: clearing dirty flags, synchronizing scroll positions, collecting absolute-positioned descendants for each ancestor, and applying position offsets

Shaping-cache hit counts do not reconcile with misses caused only by width changes

Small non-animation changes (hovering one button, typing one character, caret blinking, or scrolling) rerecord the entire page's DisplayList and recompute its full content hash

When a soft hyphen does not fit, each affected line triggers character replacement and a full paragraph relayout

Typing within a textarea paragraph and pushing a word onto the next line rebuilds the paragraph once for each subsequent visible line

Reflowing an unclamped -webkit-box with inline elements is about 4 times slower than before the fix and 1.6 times slower than an ordinary block with the same content (the measured cost of A179)

In the settings lesson, Tab can focus the button, but Enter did not trigger its click save callback.