text: Avoid quadratic table highlight remapping

[lurenjia534]

Description

A full Markdown update after highlighting search matches repeatedly scans the table from its first row for each highlighted cell. It also recomputes the target row's source endpoint for every cell in that row. Long tables and wide rows therefore cause quadratic reconciliation work on the entity/UI update path.

Collect table row endpoints once in LeafRemap::new and find each cell's row by binary search. Preserve whole-row highlight invalidation, conservative handling of missing source positions, and nested/cross-table boundaries. Append-only remapping skips the row index. Four regression tests cover these cases. The change is confined to one source file and introduces no public API or behavior limits.

Performance

Measured with the production Markdown parser, rendered-text index, public highlight setter and remap. Every x is highlighted, then the final cell of the first data row changes to y. Timings include LeafRemap construction and RangeHighlightFrame::remap.

Table shape (including header)	Highlighted cells	Before	After
32,001 rows × 1 column	32,001	146.280 ms	2.392 ms
64,001 rows × 1 column	64,001	590.392 ms	5.711 ms
100,001 rows × 1 column	100,001	1,467.864 ms	10.104 ms
3 rows × 4,096 columns	12,288	180.165 ms	0.866 ms

Linux x86_64 / Ryzen 9 7950X / Rust 1.98.0; gpui-base uses opt-level=3, with dependencies using the repository's dev profile. Each result is the median of five total-time samples after warm-up. Parsing, search and highlight installation are excluded. These are reconciliation measurements, not GUI frame latency or a full release build.

Screenshot

The screenshots show the actual Rust benchmark output in native terminal windows.

Before	After

Benchmark source, raw results and reproduction instructions are stored separately from the fix so the PR diff contains only the implementation and regression tests.

How to Test

cargo test --locked -p gpui-base --lib
cargo fmt --all -- --check
cargo clippy --locked -p gpui-base --all-targets -- -D warnings
git diff --check

All checks passed; the final library suite reports 1,152 passed, 0 failed. The new tests cover empty cells and missing endpoints, nested tables and table boundaries, long/wide table row invalidation, and the append fast path.

A temporary integration test also passed through public set_text and the background parsed-result commit for a 2,002-highlight table. It verified preserved header highlights/reveal and cleared highlights in the edited and following rows. This instrumentation was removed after validation; its source is included with the linked evidence.

env -u WAYLAND_DISPLAY cargo run --locked -p example-markdown built and displayed on Linux. This was a startup smoke check; full GUI stress testing and macOS/Windows performance testing were not completed.

AI Assistance

AI assisted with the patch, code review, performance reproduction, regression validation and PR preparation.

Checklist

• I have read the contributing guide and kept this PR focused on one problem.
• Reviewed the implementation and AI-assisted code against the test evidence.
• Passed cargo run for related Story tests (Markdown example startup smoke check only).
• Tested macOS, Windows and Linux performance (Linux reconciliation measurements completed; no platform-specific code changes).

[huacnlee]

Thank you for the deep performance investigation and optimization, backed by clear benchmarks and regression tests. This kind of careful performance work is very important to GPUI Kit. Approved; Security: PASS for the affected scope.