In https://github.com/lightpanda-io/browser/pull/1774 we started to track Ranges
in the page in order to correctly make them "live". But, without correct
lifetime, they would continue to be "live" even if out of scope in JS.
This commit adds finalizers to Range via reference counting similar to Events.
It _is_ possible for a Range to outlive its page, so we can't just remove the
range from the Page's _live_range list - the page might not be valid. This
commit gives every page an unique id and the ability to try and get the page
by id from the session. By capturing the page_id at creation-time, a Range
can defensively remove itself from the page's list. If the page is already
gone, then there's no need to do anything.
Next.js hydration references FileList as a global for feature detection.
Register a minimal stub (length=0, item()→null) so the type exists in
the global scope and the reference check doesn't throw.
Custom element callbacks aren't allowed to call document.open/close/write while
parsing.
Fixes WPT crash:
/custom-elements/throw-on-dynamic-markup-insertion-counter-reactions.html
On BSD, a listening socket isn't considered connected, so this error is
expected. Maybe we shouldn't call shutdown at all, but I guess it's safer this
way.
The bridge.finalizer resolves Page through its own module graph, which
can differ from Range.zig's import in release builds. Store a pointer
to the live_ranges list directly on AbstractRange so deinit can remove
without accessing Page fields.
Add a weak finalizer to Range that removes its linked list node from
Page._live_ranges when V8 garbage-collects the JS Range object. This
prevents the list from growing unboundedly and avoids iterating over
stale entries during mutation updates.
Move the per-range update logic from Page into AbstractRange methods
(updateForCharacterDataReplace, updateForSplitText, updateForNodeInsertion,
updateForNodeRemoval). Page now just iterates the list and delegates.
Remove redundant start_container == end_container check in insertNode —
collapsed already implies same container.
When calculating accessible names for elements without explicit labels, multiple descendant text nodes were previously concatenated directly together. This adds a space between distinct text node contents to prevent words from sticking together.
Previously, semantic_tree_text hardcoded prune = false, which bypassed the structural node filters and allowed empty none nodes to pollute the root of the text dump.
This prevents token bloat in JSON/text dumps and ensures that StaticText leaf nodes are not incorrectly pruned when structural containers (like none, table) hoist their text.
This fixes a bug where only the first text node was being returned, causing fragmented text nodes (e.g. <span>Sub</span><span>mit</span>) to be missing their trailing text.
