4.12.36

🐛 Bug Fix

• Lists / enter: 'br' mode: converting a list type (ul ↔ ol, or applying a class/style that forces the REPLACE path) destroyed the list — unwrapList extracted the li content into a DocumentFragment that was dissolved on insertion, so wrapList received an empty detached node and the new <ol>/<ul> was never inserted into the document; the content was left as bare inline nodes. In the REPLACE path the li now stays in the document and is re-wrapped directly.
• commitStyle / attributes: a number/boolean attribute value (e.g. commitStyle({ element: 'a', attributes: { tabindex: 2 } })) could never be toggled off — the attr() getter returns a string, so the strict comparison against the raw value always failed and every re-apply took the CHANGE branch instead of UNSET. The value is normalized to a string before comparison now.
• Selection: cursorInTheEdge() (and its cursorOnTheLeft/cursorOnTheRight wrappers) threw IndexSizeError when the document had no selection ranges — getRangeAt(0) was called without checking rangeCount. It returns null now, as documented.
• Selection: eachSelection() threw TypeError when the selection root was an empty editor (childNodes[-1] produced an undefined start node).
• Dom.isFragment: fragments belonging to an inert document (template.content, DOMParser output) were not recognized because of a stale defaultView requirement — same class of problem already fixed for Dom.isElement earlier.
• Dom.between: when end was an ancestor of start, the ascent skipped over it and the callback kept firing for nodes far outside the intended range, up to the end of the tree.
• Dom.replaceTemporaryFromString: the regular expression hardcoded the data-jodit-temp attribute name instead of using the TEMP_ATTR constant and did not match a temporary element whose marker attribute has no value (<span data-jodit-temp>).
• LazyWalker: stopping a walk did not really cancel it — stop() called cancelIdleCallback with an id that was never assigned (dead code left from the pre-scheduler.postTask implementation) and the scheduler's AbortController was never aborted. Also, setWork() called before the first chunk had started did not cancel the previously scheduled pass. Repeated setWork() calls (the clean-html plugin does this on every change) piled up concurrent loops pumping the same generator, defeating the chunked-walk throttling. The pending task is now aborted in stop().
• LazyWalker: the affect flag was not reset when a pass was interrupted via break(), so the next pass reported end(affect = true) even if it had not changed anything (a false synchronizeValues trigger for clean-html). It also processed timeoutChunkSize + 1 nodes per chunk instead of the configured size.

🏠 Internal

• Dom: the Dom.replace(elm, 'p') overload without a create instance is removed from the type surface — it always failed at runtime on an assertion (Need create instance for new tag); create is now required whenever the replacement is defined by a tag name or an HTML string. Dom.isList is honestly typed as HTMLUListElement | HTMLOListElement, Dom.isComment accepts Nullable<Node>, isSameAttributes no longer pretends to be a type predicate, CommitStyle.isApplied returns a real boolean.
• Dom: nextGen builds its sibling stack with push + reverse instead of unshift in a loop (was O(n²) on wide sibling lists); isTag computes the upper-case tag name lazily; the temporary-element regexp is compiled once at the module level.
• Docs: fixed the misleading/inverted JSDoc for Dom.between, Dom.up (root-check asymmetry), Dom.safeInsertNode, LazyWalker options and events, cursorInTheEdgeOfString, isSuitElement (strictStyle description was inverted), the LazyWalker example in the DOM module README (options object, not a number), and the Dom.replace example (missing create argument).
• Tests: new coverage for Dom.between, Dom.isFragment (inert documents), Dom.replace, Dom.replaceTemporaryFromString, wide sibling traversal order in Dom.find, LazyWalker restart/affect-flag reset, cursorInTheEdge without a selection, eachSelection on an empty editor, ul → ol conversion in enter: 'br' mode, and toggling a numeric attribute via commitStyle.
• Tests / hermetic images: the resize, image-editor and image-properties karma tests that wait for a real image load no longer fetch https://xdsoft.net/jodit/files/artio.jpg/th.jpg from the live server — they use the byte-identical tests/artio.jpg already served by karma, so a slow or unreachable xdsoft.net can't time these tests out anymore (~17 tests were flaking on degraded networks). The onLoadImage test helper now also rejects with a clear failed to load "<src>" message on an image error event instead of hanging until the mocha timeout.