test(loro.js): add a multi-peer differential fuzz harness

[zxch3n]

Update (merged main at 238988f5, with #1131): a second fixed run adds 120 seeds with nested child containers and snapshot imports, all passing on main now; both runs take about 5.5 s locally. Seeds 297, 315, 344, 422, 452, and 559 of 0–599 with nested children still fail the event oracle on main; #1150 fixes all six.

Update (merged main at c0ff1255): #1126 and #1136 fix all seven seeds that were pending, so the default run now covers all 120 seeds, and the full-snapshot oracle is on as well (about 3 s locally; the timeout is 120 s). FUZZ_WASM=1 now uses the shared tests/support/rust-reference.ts loader from #1136, so LORO_WASM_NODEJS and the CI reference build apply. Lockstep with Rust is only meaningful once #1145 (absent-key map deletes, zero counter increments) and #1143 (checkout of the current version) land. On main most seeds diverge at the first such call; with #1145, 49 of 50 seeds match step by step. The context doc lists what is fixed, what is in PRs, and what is still open (tree cycles and tree event order).

Summary

Adds a randomized differential harness for loro.js and keeps a CI-sized fixed seed set in the test suite.

• loro-js/tests/fuzz/differential.ts: 2–4 peers make random edits (Map, List, Text, Tree, Counter, MovableList, marks, child containers), exchange updates (update, full, per-change blobs in shuffled order via import or importBatch, optionally snapshots), and check out random versions.
• Oracles: every peer equals a one-shot replay and a shuffled per-change replay; checkout(F) equals a fresh import of exactly the changes in F; a full snapshot and shallow snapshots equal the full-history document, including checkouts into the retained range; an event mirror keyed by container ID must reproduce toJSON() after every commit, import, and checkout. The mirror follows the Rust fuzzer's ContainerTracker rules.
• Optional Rust reference (FUZZ_WASM=1): each peer gets a loro-crdt twin, every concrete edit/sync/checkout runs on both, and the first diverging action is reported. It also compares a Rust replay, Rust checkouts, both shallow-snapshot directions, and the chosen shallow root.
• Failures are grouped by signature, minimized by delta debugging over the action list, and reported as a trace of loro.js API calls.
• context/loro-js-differential-fuzz.md: how to run it, the oracle semantics, and the divergences found on main, linked from the root AGENTS.md.

Rationale

The runtime had no multi-peer oracle beyond hand-written cases. This harness found the bugs fixed in #1139, #1140, and the other fix/loro-js-* PRs from this session. It also found the still-open Tree, nested-container event, and shallow classes listed in the context doc. The default run needs no WASM build, so it can run with the rest of pnpm test. It overlaps little with #1132, which pairs every peer with Rust for MovableList only.

Validation

• Default run: 120 seeds of Map/List/Text/Counter and 120 seeds with nested child containers and snapshot imports, with the replay, checkout, full-snapshot, and event oracles, about 5.5 s; all pass on main at 238988f5.
• FUZZ_VALIDATE_HARNESS=1 with the WASM reference found no case where Rust's own events disagree with the mirror: 300 Map/List/Text/Counter seeds with child containers and 100 CORE_FEATURES seeds (which include trees). A tree-only validation run could not finish because loro.js loops forever on a tree cycle created by concurrent moves (listed in the context doc).
• pnpm --dir loro-js format:check, lint, typecheck, test (249 passed, 1 skipped manual block), build.
• No changeset: tests and docs only; the published package does not change.

🤖 Generated with Claude Code

[github-actions]

WASM Size Report

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