perf: improve text checkout scalability

[lodyai]

Summary

• Add a text checkout benchmark with profile counters for frontier preparation, VV conversion, diff calculation, richtext tracker work, state apply, event emit, and future-sibling scanning.
• Reduce per-change VersionVector work by passing a lightweight causal version view through diff calculation and teaching the richtext tracker to consume it directly.
• Use a forward diff calculator for comparable checkout-to-latest paths so linear/import-greater updates avoid persistent CRDT checkout tracking.
• Add plain-text no-style apply fast paths and conservative same-deps / same-parent fast paths for high-collaboration text checkout.
• Keep FastUpdates/FastSnapshot text imports on the fast path by avoiding the state-apply rollback guard for ordinary text sync, while still forcing rollback for JSON schema imports from external input.

Benchmarks

Text checkout improvements versus main:

• 1000 peer wide-causal checkout: about 5.13ms -> 1.61ms, diff_calc about 4.90ms -> 1.39ms, tracker checkout about 3.47ms -> 37.6us.
• 300 peer same-position checkout: about 4.93ms -> 1.78ms, frontier_prepare about 3.04ms -> 37.8us.
• 1000 peer same-position checkout after frontier fast path: about 16.6ms. Future-sibling scan profile after same-parent fast path: about 1.83ms -> 575us.
• checkout-to-latest linear smoke: about 65us with richtext tracker checkout/diff calls at 0.

Full bench rerun notes:

• Re-ran all benches against the merge-base using separate target dirs to avoid cross-branch artifact contamination.
• The text checkout benchmarks are the intended wins, generally in the 3x-9x range depending on the scenario.
• During the full run, three non-checkout text sync/import benches initially regressed by more than 50% because text imports were also forcing state-apply rollback. After narrowing rollback to list/tree for ordinary binary imports and forcing it only for JSON schema imports, the regressions recovered:
  • encode_with_sync/update: regressed to about 88ms, now about 46.0ms; base was about 43-45ms.
  • refactored direct_apply/B4DirectSync: regressed to about 6.7s, now about 3.69s; base was about 3.6s.
  • refactored-sync/B4Parallel: regressed to about 89.7ms, now about 54.0ms; base was about 51.6ms.

Validation

• cargo check -p loro-internal --features test_utils,jsonpath
• cargo test -p loro-internal --features test_utils,jsonpath import --lib
• cargo test -p loro --test panic_test import_json_updates_with_text_insert_out_of_bounds_should_error_without_mutating_doc
• pnpm check
• pnpm test
• CARGO_TARGET_DIR=/tmp/loro-final-bench cargo bench -p loro-internal --features test_utils,jsonpath --bench encode encode_with_sync/update -- --save-baseline final-encode
• CARGO_TARGET_DIR=/tmp/loro-final-bench cargo bench -p loro-internal --features test_utils,jsonpath --bench text_r B4DirectSync -- --save-baseline final-direct
• CARGO_TARGET_DIR=/tmp/loro-final-bench cargo bench -p loro-internal --features test_utils,jsonpath --bench text_r B4Parallel -- --save-baseline final-parallel

Notes

• B4Parallel filtering also matches DecodeUpdates B4Parallel; after the target bench reported 54.0ms, the remaining matched bench was terminated.
• Long-running libFuzzer targets were not run.

[lodyai]

Benchmark / validation update after the TMP_PLAN cleanup.

Fixed-size checkout bench command shape:

LORO_TEXT_CHECKOUT_PROFILE=1 LORO_TEXT_CHECKOUT_PEERS=1000 LORO_TEXT_CHECKOUT_BASE_LEN=1024 LORO_TEXT_CHECKOUT_CHANGES=1000 cargo bench -p loro-internal --features test_utils --bench text_checkout -- <case> --warm-up-time 0.05 --measurement-time 0.1 --sample-size 10

Results:

case	criterion median-ish range	avg_total	main profile notes
plain/wide-causal-peer-checkout/1001	1.37-1.67 ms	1.46 ms	diff_calc 1.27 ms; richtext_tracker_checkout 33 us; future_scan 0
plain/same-position-peer-checkout/1001	12.0-16.9 ms	15.89 ms	diff_calc 15.15 ms; future_scan 0.57 ms; avg_future_scan_visited 383, max 999
code/checkout-to-latest-linear/1001	515-524 us	521 us	diff_calc 358 us; state_apply 160 us; richtext tracker 0
rich/overlap-mark-peer-checkout/subscribed/1001	11.2-17.1 ms	15.20 ms	state_apply 7.66 ms; diff_calc 6.63 ms; emit_events 72 us; future_scan 31 us

Phase 5 cost note: after the same-parent fast path, future sibling scanning is no longer the dominant cost in the measured 1000-peer cases. It is still visible in the worst same-position case (~0.57 ms, about 3-4% of total), but the remaining slow path is mostly replay/diff_calc. For subscribed rich overlap marks, future_scan is negligible (~31 us), while state apply and diff calc dominate.

Additional validation now passed:

cargo fuzz run text-update -- -max_total_time=60
cargo fuzz run all -- -max_total_time=60

Both fuzz targets completed their 60s runs without crashes. The fuzz lockfile refresh was discarded because it was generated by the local run and is unrelated to this PR.

[lodyai]

Follow-up phase 4 update after continuing the 1-4 plan.

Added commit 377afd4e (perf: batch rich text style event deltas):

• Batch adjacent, non-overlapping retain-only rich text style event deltas before composing them.
• Preserve the original compose order by flushing when a delta overlaps or is not retain-only.
• This keeps the optimization local to event conversion and does not add checkout/cache state.

Validation passed after this commit:

cargo check -p loro-internal --features test_utils
cargo test -p loro-internal richtext --features test_utils
cargo test -p loro-internal checkout --features test_utils
cargo test -p loro-internal import --features test_utils
cargo test -p fuzz random_fuzz_1s -- --nocapture
git diff --check

Affected bench rerun:

LORO_TEXT_CHECKOUT_PROFILE=1 LORO_TEXT_CHECKOUT_PEERS=1000 LORO_TEXT_CHECKOUT_BASE_LEN=1024 LORO_TEXT_CHECKOUT_CHANGES=1000 cargo bench -p loro-internal --features test_utils --bench text_checkout -- rich/overlap-mark-peer-checkout/subscribed --warm-up-time 0.05 --measurement-time 0.1 --sample-size 10

Result: Criterion range 8.7449 ms - 13.709 ms, avg_total=12.692376ms, avg_state_apply=6.503316ms, avg_diff_calc=5.456606ms, avg_emit_events=50.446us, avg_richtext_insert_future_scan=27.024us. The sample is noisy (p = 0.17, no statistically significant change), but the profile is directionally better than the previous run (avg_total=15.199131ms, avg_state_apply=7.66148ms).

[zxch3n]

Fuzz validation for the text checkout performance PR:

• cargo fuzz run text-update -- -max_total_time=1200
  • Passed: 317,498 runs in 1201s
  • Final: cov: 4634, ft: 18712, corp: 1220/255Kb, rss: 507Mb
• cargo fuzz run local_events -- -max_total_time=1200
  • Passed: 510,962 runs in 1201s
  • Final: cov: 15694, ft: 59080, corp: 2001/428Kb, rss: 566Mb
• cargo fuzz run all -- -max_total_time=1200
  • Passed: 52,862 runs in 1201s
  • Final: cov: 26017, ft: 74208, corp: 625/22Kb, rss: 471Mb

No crashes, panics, sanitizer failures, or reproducer artifacts were reported. Workspace was clean after restoring the fuzz-generated crates/fuzz/fuzz/Cargo.lock refresh.

[github-actions]

WASM Size Report

• Original size: 3111.82 KB
• Gzipped size: 999.85 KB
• Brotli size: 696.54 KB