[zxch3n]

Summary

Found while auditing loro.js against the recent Rust snapshot fixes: some probes exported a snapshot after a checkout.

On main, export({ mode: "snapshot" }) on a detached loro.js document writes the checked-out state, the checked-out version vector, and frontiers next to the full history. The resulting blob is inconsistent:

• The importer reports the old state and old frontiers as latest, for example {"m":{"a":1}} with frontiers 0@1 while the history goes up to 1@1.
• checkout on the importer throws snapshot version does not match its history.
• Updates imported while detached are missing from the exported state.

Rust's encode_snapshot_inner checks out the latest version, encodes, and checks the old version back out. loro.js now does the same: it encodes the latest state, writes the oplog version vector and frontiers, and then restores the checkout. The exporting document stays detached at the same version.

Validation

• New loro-js/tests/detached-snapshot.test.ts: the export keeps the checkout; the importer sees the latest value and frontiers and can check out the old version; updates imported while detached are included. Both tests fail on main.
• Cross-runtime check: Rust main imports a loro.js snapshot exported while detached and gets the latest value {"m":{"a":2},"t":"x"} and frontiers 2@1.
• pnpm --dir loro-js test (18 files, 242 tests), typecheck, lint, format:check all pass.
• The non-detached path is unchanged. B4 (bench:b4 259778 3) snapshot export is 118.9 ms here vs 112.2 ms on main, with overlapping samples (109–123 vs 107–122 ms) and identical output (309,780 bytes). Only a detached export pays for rebuilding the latest state and restoring the checkout, which is O(history), as in Rust's two checkouts.

The test history avoids a list delete, because retreating a delete on snapshot-imported state is a separate bug fixed in #1126.

🤖 Generated with Claude Code

[github-actions]

WASM Size Report

• Original size: 3200.65 KB
• Gzipped size: 1058.59 KB
• Brotli size: 741.14 KB

[zxch3n]

Addressed the adversarial review in 52710c1.

P1: the lazy snapshot branch returned early. #encodeSnapshot now computes latestVersion and detachedState first. The lazy (deferred) branch runs only when the state is already at the latest version, and it writes latestVersion and the oplog frontiers instead of this.version() / #frontiersCodec(). Your repro (import snapshot → detach() → import x=2 → export snapshot) is now a regression test in tests/detached-snapshot.test.ts. The importer gets {m:{x:2}} with frontier 1@1, and checkout of the latest version succeeds. The exporter stays detached at {m:{x:1}}. Rust main imports the same blob as {"m":{"x":2}} / 1@1, and checkout succeeds. The test fails on the old head 429d6dc.

No more double replay. The new #latestStateStore() uses the same incremental transition as diff(): go forward to the latest version, build the store, and go back to the checkout, which is O(version delta). If the transition is refused, or the state is still lazily encoded (which #buildStateStore cannot enumerate), it replays once on an isolated staging doc (forkAt(oplogFrontiers())). The exporter's containers and handles are never rebuilt.

10k map keys, 100 edits after the checkout, median of 21 runs per process, 3 alternating rounds (ms):

	main (wrong output when detached)	old head 429d6dc	52710c1
attached	18.46 / 18.81 / 18.66	18.54 / 18.19 / 18.49	18.04 / 18.17 / 19.17
detached	18.14 / 18.14 / 17.83	34.47 / 33.83 / 35.12	18.52 / 18.78 / 19.54

The export test now also deletes text, checks that a held LoroText handle keeps its checkout value, and that checkoutToLatest() still works after the export. The importer-side historical checkout across that delete depends on #1126. On the combined tree (main + #1126→#1127→#1128 + #1130 + #1131), loro.js and the Rust main nodejs WASM build both import the detached export at the latest value and check out the old value correctly.

Checks: pnpm --dir loro-js test (18 files, 243 tests), typecheck, lint, and format:check all pass. The combined tree passes 22 files, 284 tests. The changeset text is updated.