Service Worker and Cache gain WPT baselines
The DOM corpus reached 9,734 passes with 135 failures remaining. Initial Service Worker fetch and CacheStorage baselines passed, alongside CSS hardening and Android rendering work.
Overview
The js-dom track finished 76 rounds this week (R118–R193), moving the full WPT DOM corpus from 9,727 to 9,734 passes with 135 failures left. The work spanned query scoping, cross-realm adoption and listener semantics, live Range getters, DOMParser mutation surfaces, MutationObserver records and namespace attribute metadata.
The CSS parser and style system took a long hardening tail: non-finite values rejected, keywords parsed case-insensitively, dozens of commits in total, with a layout length-resolution batch and a layout performance series (R3429–R3447) alongside. Service Workers landed its first WPT baselines: 6 fetch-interception cases and 10 CacheStorage cases, every subtest passing; the sibling CacheStorage window baseline reached 32 cases / 429 subtests at full pass.
The Android renderer gained real browsing capability: frames forwarded to the compositor, scroll gestures, network proxying, back navigation and a persistent profile. CI came back 9/10 green, the regression guard found no new regressions, and the perf gate passed a full 113-metric rerun.
Completed
js-dom M4, rounds R118–R193: querySelector CSS escape decoding, spec validation for removeChild/replaceChild/insertAdjacentElement/importNode, the CharacterData and cloneNode method families, name-validation regexes, getRootNode with composed shadow roots, the native overlay event family.
Later rounds added the L2-d3c compound gate for document queries, ns-prefix selectors at any segment, :lang/:target element-query context, fragment query scoping, live Range offsets and commonAncestorContainer, a real normalize and textContent, cross-realm adopt/listener/creation channels, DOMParser mutable factories, namespace attribute metadata (qualified names, clone ns fixup, prefix derivation), MutationObserver document-target observation, DocumentFragment.getElementById and strict XML names for processing-instruction targets.
Corpus went 9,727P/141F → 9,734P/135F, net +7P/−6F, zero regressions.
CSS parser and style-system hardening: non-finite values rejected in colors, transforms, angles, lengths, media queries, grid tracks, form controls and replaced-element attributes; keyword parsing made case-insensitive (fit-content, grid span, var/env, supports, clip-path, border-image, math functions and more); background layers split by top-level commas; empty gradient stops, unclosed transform functions and zero grid spans rejected; quoted font-family names split correctly.
A layout length-resolution batch covered table columns, declared widths, abspos insets and intrinsic lengths; the layout performance series R3429–R3447 added snapshot flags, IFC metric reuse, fast hashing and a sparsified font context.
Service Worker fetch and Cache WPT baselines: fetch runner 6 cases / 10 subtests all pass (request end-to-end, async respondWith, network error, respondWith argument, iso-latin1 header, add-async); CacheStorage serviceworker runner 10 cases / 154 subtests all pass; respondWith results survive handler throws; Cache.delete and CacheStorage.delete/has/keys share one typed host bridge.
The sibling storage-cache-api goal took the CacheStorage window baseline to 32 cases / 429 subtests at full pass, added per-origin CacheStorage persistence to disk (temp-file + sync + atomic replace) and storageBuckets name-prefix isolation.
Android renderer: WSL build with V8 toolchain provisioning, renderer frames forwarded to the compositor, page scroll gestures, network request proxying, external HTTP intents, back navigation, persistent profile through browser-shell.
CI and measurement: 9/10 jobs green on main; the one failure is the pre-existing windows-aarch64 clippy environment issue, tolerated by continue-on-error. The full CPU reftest passed. ZRG-2026-08-23-01 found no new regressions: the perf-gate rerun cleared all 113 metrics with NEW=0 and compositor peak memory stayed in budget for a fourth consecutive round. Perf baselines and trends were auto-tightened and committed; the WPT trend baseline was refreshed by the weekly cron.
In progress
js-dom moves to R194 next: rescan the remaining 135 failures (ShadowRoot Range operations, attach-shadow-realm and svg-template querySelector are the new top clusters), keep the M2/M6 items going (de-stringing high-level API bridges, completing native dom_bindings) and revisit the customElements multi-registry deep item. Service Workers keeps widening the fetch/cache WPT surface, with M3 clients follow-up (popup/auxiliary browsing contexts) coming after. storage-cache-api continues past its persistence first slice.
The rendering-compat stream is doing the approved oracle recapture and parity-threshold relaxation at its own cadence, with grid-template shorthand grammar as the current rally focus. One item waits for a physical-GPU machine: split-process idle memory totals 270.5 MiB against a 256 MiB budget, hanging for the ninth consecutive round on Xvfb/llvmpipe.
Sources
Engineering guides
How ZeroWeb tests AI-autonomous software engineering
A browser gives AI development a long-running test case: architecture, code, debugging and verification must work together.
ZeroWeb Rust browser architecture: from request to pixels
Follow the page pipeline and the boundaries between the browser application, embeddable WebView, engine and rendering infrastructure.
Getting started with ZeroWebView and ZeroBrowser
Prepare a local checkout, launch the desktop browser and inspect the WebView API before embedding the experimental engine.
How to read ZeroWeb WPT and performance results
Understand self-source reftests, assertion-based suites, changing corpus sizes and p95 benchmarks before comparing results.