fix(loro.js): treat containers under deleted ancestors as deleted

[zxch3n]

Summary

loro.js decided whether a container was deleted from its own parent binding only, and a tree node from its own deleted flag, so anything below a deleted container or tree node counted as alive. loro-crdt follows the whole chain. Differences on main, checked against a WASM build of main (c0ff1255):

After	Call	main	loro-crdt / this PR
root.delete("m") where m holds a text	text.isDeleted()	false	true
same	doc.getPathToContainer(text.id)	["root","m","t"]	undefined
tree.delete(parent)	tree.getNodes()	includes the children	alive nodes only
same	tree.isNodeDeleted(child), child.isDeleted(), grandchild.data.isDeleted()	false	true
any	tree.nodes()	alive nodes	every node, including deleted ("Get all tree nodes of the forest, including deleted nodes")

Rationale

Document._isContainerUnreachable walks the parent chain, and for tree metadata the tree ancestors (LoroTree._isNodeHidden, bounded by the node count so a malformed cycle cannot loop). It backs isDeleted(), getPathToContainer(), and getCursorPos() only. getNodes() now lists alive nodes breadth-first from the roots, then (with withDeleted) the deleted subtrees, as Rust's get_nodes_under does.

Changed after #1131 merged: _isContainerDeleted stays main's one-level check. It is used by getContainerById and by #1131's #isReachable, which already walks ancestors; walking them again per level made deep trees quadratic (tree-diff-order.test.ts). The earlier guard that rejected local edits to deleted containers is dropped: it threw on a redo that main (and #1131's diff-attached-children.test.ts) accepts.

Known limitations (unchanged from main)

• loro-crdt rejects local edits to a deleted container (Transaction::apply_local_op); loro.js still records them.
• For a directly deleted node, Rust's parent() returns its internal deleted-root node and index() returns undefined; loro.js returns undefined and -1.

Validation

• New loro-js/tests/deleted-containers.test.ts (map descendants, list children, tree subtrees including revival); all three fail on main.
• pnpm -C loro-js check with LORO_REQUIRE_WASM_REFERENCE=1 and a Rust reference built from main: 29 files, 482 passed, 1 skipped.
• The local edit, import, and diff paths are main's code again; bench:complexity for the unchanged getNodes/nodes change (7 alternating rounds vs main, loadavg 10–14): 284 items, none both >1.5× and >0.3 ms slower.

🤖 Generated with Claude Code

[github-actions]

WASM Size Report

• Original size: 3277.59 KB
• Gzipped size: 1089.33 KB
• Brotli size: 762.83 KB

[zxch3n]

Merged the latest main (c0ff1255) with no conflicts. pnpm -C loro-js check with LORO_REQUIRE_WASM_REFERENCE=1 passes (26 files, 417 tests), including #1136's Rust differential suite, so the new "container is deleted" guard does not trip on its undo/revertTo paths.

Still needed on main: edits to a container under a deleted ancestor are accepted (Rust rejects them), isDeleted()/getPathToContainer() ignore deleted ancestors, and getNodes() lists nodes under a deleted node. All liveness probes now match loro-crdt.

Known limitation (unchanged, not a data issue): for a directly deleted node, Rust's parent() returns its internal deleted-root node and index() returns undefined; loro.js returns undefined and -1. #1131's #isTreeRecordHidden and this PR's LoroTree._isNodeHidden compute the same thing and can be merged in a follow-up once both land.

[zxch3n]

Merged main (with #1131) and reworked in 03dc8dc: the ancestor walk now backs only isDeleted(), getPathToContainer() and getCursorPos(); _isContainerDeleted is main's again, so #1131's reachability walk stays linear in depth. The local-edit guard is dropped (it broke a redo that main allows) and listed under known limitations. pnpm -C loro-js check with the Rust reference: 482 passed, 1 skipped.

[zxch3n]

Merged main and fixed the blocking issue in c783b88: getNodes() and _subtreeRecords append records in a loop instead of push(...list), so a node with more children than the argument limit (or that many deleted nodes with withDeleted) no longer throws a RangeError. The limit is about 125k on Node's main thread but 400k–500k in vitest's worker threads, so reproducing it in CI would need about 500k nodes; the new test instead checks the widest push() during getNodes() at 2,000 children (2,000 before the fix, 1 after). A one-off 600k-node run passes locally with the fix. pnpm -C loro-js check with the Rust reference: 32 files, 497 passed, 2 skipped.