[lexical][lexical-selection][lexical-list] Bug Fix: Skip named-slot values in block converters, diagnose replace() on a slot value, and delete slot-bearing decorator hosts at the selection boundary

[etrepum]

Description

Follow-up to #8901 (fix for #8894), which made $setBlocksType skip named slot values. The same crash class exists in every other selection-driven block converter that resolves the current block and replaces it through the tree API, and a direct replace() of a slot value still fails with the unhelpful generic "Expected node %s to have a parent" invariant. This audits and fixes the rest, applying the same rule: a slot value's assignment is managed by the node or extension that owns the slot, so generic converters leave it as-is. The eligibility checks test the slot up-link ($getSlotHost) rather than parentless-ness, so a detached node is never conflated with a slot value and keeps failing loudly.

• lexical: LexicalNode.replace() on a slot value now throws a slot-specific invariant before any mutation, including the keys and types of the value and its host for debugging (replace: node %s (type %s) is slotted into host %s (type %s); … Use $setSlot on its host to assign a replacement.), mirroring the existing $removeFromParent guard for remove() on a slotted node.
• @lexical/selection: the deprecated $wrapNodes no-ops when the selection lives in a bare-block slot frame (it previously threw in the sibling walk); a shadow-root slot value still wraps its own subtree normally.
• @lexical/list: $insertList skips slot values in its empty-block branch, its empty-list-item branch, and its walk-up list-type-change branch; $removeList skips a slotted top-level list (they previously threw or vacated the slot).
• @lexical/markdown: audited, already safe — the element/multiline shortcut runners bail unless the block's parent is a root or shadow root, which excludes slot values.

The rule is documented in the named-slots concepts page.

The audit also turned up the inverse problem: deleting a slot-bearing host from outside. Backspace with the caret at the boundary after a slot-bearing DecoratorNode host (e.g. the playground PullQuote) was a silent no-op — deleteCharacter special-cased slot-bearing decorators to never be removed via backspace, leaving them undeletable from the keyboard. That special case is removed, so a slot-bearing decorator behaves like any other block decorator at the boundary: backspace removes it as a whole unit (slots included), and from an adjacent empty paragraph it drops the paragraph and node-selects the host so the next backspace deletes it. The ElementNode-host merge guard is intentionally unchanged: a cross-block merge would keep the host's children while silently dropping its slots, so backspace at the start of a slot-bearing host's child still leaves the host in place.

Closes #8904

Test plan

Before

$insertList, $removeList, and $wrapNodes threw "Expected node %s to have a parent" (or silently vacated a slot) when the selection was inside a bare-block named-slot value or a slotted list; replace() on a slot value threw the same generic invariant. Backspace after a slot-bearing decorator host did nothing; the host could not be deleted from the keyboard.

After

The converters leave slot values as-is and replace() on a slot value throws a diagnostic pointing at $setSlot. Backspace removes a slot-bearing decorator host at the boundary (or node-selects it from an adjacent empty paragraph, matching normal block decorator behavior). New unit tests in LexicalSetBlocksTypeSlots.test.ts and LexicalFormatListSlots.test.ts, plus replace() and boundary-deletion tests in LexicalSlot.test.ts; the existing test that a slot-bearing ElementNode host is not merged away still passes.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
lexical	Ready	Preview	Aug 2, 2026 4:40pm
lexical-playground	Ready	Preview	Aug 2, 2026 4:40pm

[potatowagon]

Ship-worthy. ✅ Checked out the PR, ran the tests (136/136 pass, incl. the 101-test LexicalSlot suite), and traced the trickiest claim against source. No blocking issues — clean, well-scoped finish to the named-slots audit #8901 started.

✅ What's right (verified)

1. The deleteCharacter special-case removal is correct. Traced the fall-through: removing the slot-bearing-decorator branch drops it into the generic $isDecoratorNode branch, which does exactly what's described — empty adjacent paragraph → node-select the host; otherwise → caret.origin.remove() as a unit. Slots ride along because they hang off the host node itself.
2. A detached node is NOT conflated with a slot value. The eligibility gates test $getSlotHost() !== null (the __slotHost up-link), so a truly parentless/detached node returns null and keeps failing loudly on the generic invariant. That's the subtle correctness point and it holds.
3. The ElementNode-host merge guard is intentionally left in place. The asymmetry is deliberate and well-reasoned: a cross-block merge keeps children but silently drops slots, so backspace-at-start still refuses; the decorator case removes the whole unit. Both paths have tests.
4. formatList.ts is disciplined. All three $insertList branches (empty-block, empty-list-item, walk-up) plus $removeList gate on $getSlotHost(...) === null with a comment at each site.
5. Tests genuinely test the claim. The replace() test asserts "nothing was mutated by the failed replace" and matches the diagnostic; the two boundary-deletion tests cover both the remove-as-unit and node-select-from-empty-paragraph paths.

🟡 Minor (non-blocking)

1. markdown was only audited, not tested. The description says @lexical/markdown is already safe because the block runners bail unless the parent is a root/shadow root. That's a real invariant, but there's no regression test pinning it — if someone later loosens that parent check, nothing fails. A one-line "markdown shortcut is a no-op in a slot value" test would lock it in.

🔵 Nits (cosmetic)

2. The $runInEditor → runInEditor rename (8/8 in LexicalIsAtEdgeOfElement.test.ts) is unrelated lint-conformance churn bundled into a bug-fix PR. Harmless (grep confirms no leftover refs), just diff noise.

🔎 Uncaught cases considered — all OK

• Shadow-root slot value (vs bare-block): $wrapNodes correctly still wraps its own subtree via the !$isRootOrShadowRoot(slotFrame) check. ✓
• Selection crossing a slot boundary: a selection never crosses a slot boundary, so checking the anchor covers the whole selection. ✓
• Isolated decorator: still short-circuits before the removed branch, unaffected. ✓
• replace() with a text node onto root: errorOnInsertTextNodeOnRoot still runs first, before the new slot guard. ✓