fix: skip redundant text marks when the range spans multiple style ranges

[frioux]

Summary

TextHandler::mark_with_txn already tries to skip marks that would not change anything, but the check relies on StyleRangeMap::get_styles_of_range, which returns None whenever the queried range spans more than one style-range leaf. On any document whose style ranges are fragmented — i.e. any document that has styles — the dedup silently never fires, and re-asserting a byte-identical mark records a new StyleStart/StyleEnd pair every time.

That matters because each redundant mark op leaves a pair of style anchors in the text container's state forever: they survive (shallow) snapshots, are never consolidated even when fully superseded, and every styled read (get_richtext_value / toDelta) walks all of them, uncached. A well-meaning caller that re-asserts mark state (as loro-prosemirror updateLoroText does on every change — see loro-dev/loro-prosemirror#82) can permanently and invisibly ruin a container. Measured with the included perf guard on a 724-char paragraph:

redundant re-marks	styled read (before)	styled read (after)
0	5.7 µs	4.4 µs
1000	4,156 µs (~730×)	3.1 µs
2000	21,586 µs	4.1 µs
4000	87,136 µs	3.2 µs

(On a real heavily-edited page we saw 16–25 ms per styled read and a 135 KB snapshot for 4 KB of visible text.)

Fix

Add StyleRangeMap::range_has_key_value, which scans the style ranges covering the queried range (mirroring the existing range_contains_key traversal) and reports whether every position already resolves the key to the given value. mark_with_txn and mark_for_detached fall back to it when the single-leaf fast path returns None, using exactly the same skip predicate as before — the LWW winner per range must equal the new value — so this only generalizes the existing dedup semantics to fragmented ranges. No entity indices, encoding, or merge semantics change; marks that alter any part of the range still apply.

This doesn't GC anchors that already accumulated (that would require entity-index-safe compaction), but it closes the path by which they accumulate from redundant marks.

Validation

• crates/loro-internal/tests/richtext.rs: three new tests — redundant mark and redundant unmark spanning style boundaries record no ops (both fail before this change), and a mark that changes part of the range still applies.
• crates/loro/tests/perf_redundant_marks.rs: #[ignore]d perf guard in the style of perf_text_insert_quadratic.rs (numbers above; run with cargo test -p loro perf_redundant_marks_do_not_degrade_styled_reads --release -- --ignored --nocapture).
• cargo test -p loro-internal (318 lib + all integration suites) and cargo test -p loro pass; cargo fmt --all; cargo clippy -p loro-internal introduces no new warnings.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KQw7nfh6KaUhQy6ZQUZwMm