fix(loro.js): stop a concurrent insert at the end of its origin's subtree

[zxch3n]

Summary

loro.js Text/List replicas could converge to different orders, and all of them could differ from loro-crdt, when an insert was concurrent with a run typed after the same character and with another concurrent element whose origin was further left.

Minimal repro (Rust gives xaaaabn for every import order; loro.js on main gives xaaaanb):

p1.getText("t").insert(0, "x"); p1.commit();
p2.import(p1.export({ mode: "update" }));
p1.getText("t").insert(1, "aaaa"); p1.commit();   // run under x
p4.getText("t").insert(0, "n"); p4.commit();      // concurrent root
p2.getText("t").insert(1, "b"); p2.commit();      // concurrent with the run

A 4-peer List variant makes one replica disagree with the others (AXBYN vs AXYBN).

Rationale

indexedFugueInsertion only looks at the direct children of originLeft. Sibling subtrees are contiguous, so the gap between two children belongs to the earlier one. After the last child, though, it set the insert position to the origin-right element. The interval can also hold concurrent elements whose origin is left of originLeft (the concurrent root above). Rust's scan in CrdtRope::insert breaks before them; loro.js jumped past them.

The fix keeps the index path. After the last child it checks whether the interval's last element descends from originLeft. If it does not, it binary-searches the boundary (descent is monotone because subtrees are contiguous). The descent test walks origin-left links but jumps over each implicit run through a new per-peer sorted list of explicit (non-consecutive) element counters in FugueOriginIndex. It costs O(explicit links · log n), like Rust's span-based scan, not one probe per scalar. The scan path already matched Rust and is unchanged.

Validation

• New tests in loro-js/tests/fugue-index.test.ts: the text repro under three import orders, a 4-peer List/Text repro, and a probe-count bound for a 16k concurrent run (fails on main). The expected values come from the same scenarios run on loro-crdt (WASM build of this main).
• Differential fuzz (harness in a separate PR), 500 seeds of Map/List/Text/Counter, replay/checkout oracles: peer-vs-replay and Rust-replay failures drop from 14 to 5 on this branch alone. The remaining 5 are the detached-import delete bug, fixed in its own PR.
• pnpm --dir loro-js format:check, lint, typecheck, test (251 passed), build.
• bench:complexity (1k–8k), 5 alternating rounds vs main, loadavg 15–16: 284 items, no item both >1.5× and >0.3 ms slower. Total 1044 ms on main, 1008 ms here.

🤖 Generated with Claude Code

[github-actions]

WASM Size Report

• Original size: 3210.11 KB
• Gzipped size: 1062.59 KB
• Brotli size: 744.13 KB

[zxch3n]

Closing: covered by #1136 (e74d64dc), which fixed the same Fugue bug (fugueSubtreeEnd stops the insert at the end of the last sibling's subtree).

Checked on main at c0ff1255: the text repro (xaaaabn) and the 4-peer List repro (AXYBN) match loro-crdt for every import order, and all 5 tests from this PR's tests/fugue-index.test.ts pass on main. An insert concurrent with a 16k/131k run under the same origin takes the same ~0.05–0.4 ms on main as on this branch, so there is no performance reason to keep the alternative implementation.