test(loro.js): add a Rust differential suite for MovableList

[zxch3n]

Summary

First PR of the stack that makes the loro.js MovableList behave like Rust. This PR only adds tests and the design; the published package does not change.

• loro-js/tests/differential/ is a randomized differential suite. Each simulated peer is a pair of docs with the same peer ID, the Rust WASM build and loro.js. The scenarios cover:

  • random multi-peer insert/delete/move/set;
  • nested child containers (Map, Text, MovableList);
  • syncs by update, updates-in-range, snapshot and multi-blob importBatch;
  • snapshot and shallow-snapshot joins;
  • checkout, diff, revertTo, and single-peer undo/redo.

  Syncs randomly cross engines: Rust imports loro.js bytes and loro.js imports Rust bytes. After every step both halves must agree on the value, oplog version, frontiers, reachable container IDs, getCreatorAt/getLastEditorAt/getLastMoverAt, and canonicalized events. diff() results must match too.

• A failing seed writes its trace to the temp dir. LORO_JS_DIFF_REPLAY=<file> replays it and LORO_JS_DIFF_MINIMIZE=1 shrinks it.

• The suite loads Rust through the shared tests/support/rust-reference.ts, like the other differential suites: crates/loro-wasm/nodejs, or LORO_WASM_NODEJS. It runs inside pnpm test, skips without a WASM build, and runs in CI through pnpm test-loro-js after release-wasm. No separate workflow or script.

• context/loro-js-movable-list.md covers:

  • the Rust MovableList contract, anchored to code;
  • why the old loro.js model diverges;
  • the replacement model, implemented in the next PR;
  • how to use the suite;
  • the divergences that are not MovableList-specific.

Findings on main

MovableList, fixed by the follow-up PRs:

1. A concurrent delete + move gives ["a","c"] in loro.js and ["a","c","B"] in Rust.
2. Rust's "move, then delete" imported into loro.js gives ["b","a"] instead of ["b"].
3. loro.js snapshots write position IDs and last-set IDs as the element ID.
4. getLastEditorAt/getLastMoverAt return the creator.
5. Checkout, diff and revertTo select different winners than Rust.
6. A same-value set is skipped (Rust records it).

The insert/delete-only profile converges on main since the Fugue fixes in #1136/#1139, so the other seven profiles are marked KNOWN_DIVERGENT. Each asserts that it does diverge, and the model PR empties the set.

The suite normalizes a few differences of form and works around two Rust issues; each is commented in harness.ts/fuzz.ts and listed in the context doc:

• Event deltas that differ only in shape, or are a no-op where Rust reports nothing, are checked by their effect.
• A revertTo whose op order Rust leaves to FxHashMap order, or whose diff() shape differs, is checked by value, and Rust's ops are replicated.
• Rust reports a wrong getLastEditorAt on shallow documents after a diff across the root, so editors are not compared there.
• Rust's updates-in-range from a shallow doc exports nothing for a span that starts before the root, so spans start at the root.

Validation

• pnpm -C loro-js check: passes (511 tests; 3 Rust-engine variants of loro.js-only tests skipped). Against a reference built from current main, the suite's 8 profiles all pass: one converges, seven diverge as expected.
• CI on this PR runs the suite in pnpm test-loro-js.

Merge order

1. This PR.
2. fix(loro.js): model MovableList like Rust (positions and LWW elements) #1166 MovableList model and Rust-compatible encoding (empties KNOWN_DIVERGENT).
3. fix(loro.js): make import and importBatch atomic #1167 atomic import/importBatch.
4. fix(loro.js): reject invalid MovableList move/set refs (port Rust #1125) #1168 MovableList move/set validation (Rust fix: reject malformed movable list move/set ops on import instead of aborting #1125).

Each later PR targets the previous PR's branch.

🤖 Generated with Claude Code

[github-actions]

WASM Size Report

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