Optimize BufferLine.copyFrom sparse map copying

[Tyriar]

Summary

Optimizes BufferLine.copyFrom for the production scroll-recycle path (BufferService.scroll → recycle().copyFrom(cachedBlankLine)) and for lines with combining characters / extended attributes.

Changes

• _copySparseMapsFrom
  • Empty-map fast path: Shared EMPTY_SPARSE_MAP sentinel (Object.create(null)) with copy-on-write for _combined. Empty _extendedAttrs uses undefined (addon-image writes this map directly without BufferLine APIs).
  • Populated-map path: Inlined _data flag scan instead of per-column _copyCellMapsFrom calls (fewer function call overheads than master).
• addons/addon-image: Fork _extendedAttrs with ??= {} before writes; optional chaining on reads (fixes shared-map regression).
• test/benchmark/BufferLine.benchmark.ts: Micro-benchmarks for scroll recycle and sparse-source copyFrom.

copyFrom is only called from BufferService.scroll when the circular buffer is full.

Micro-benchmark results

500,000 copyFrom operations per case, 5 runs, average runtime (npm run benchmark -- -s "out-test/benchmark/BufferLine.benchmark.js" out-test/benchmark/BufferLine.benchmark.js). Benchmark file from this PR; only BufferLine.ts swapped per revision. Measured at dec5452fc.

Three-way comparison (ddf2b7d vs master vs this PR)

ddf2b7d is the last merge on master before #5972 (fix(buffer): clear stale attrs and combined data in BufferLine), which replaced copyFrom's key-only map copy with an O(cols) _copySparseMapsFrom scan.

Case	ddf2b7d (pre-#5972)	master	This PR	PR vs master	PR vs ddf2b7d
Scroll recycle, cols=80	15.99 ms	62.43 ms	6.79 ms	−89.1% (9.2×)	−57.5% (2.4×)
Scroll recycle, cols=279	22.34 ms	203.22 ms	19.20 ms	−90.6% (10.6×)	−14.1% (1.2×)
Sparse source, cols=80	611.51 ms	337.19 ms	317.44 ms	−5.9% (1.1×)	−48.1% (1.9×)
Sparse source, cols=279	1539.39 ms	942.76 ms	863.35 ms	−8.4% (1.1×)	−43.9% (1.8×)

What each revision does in copyFrom

Revision	Empty-map path (yes / scroll recycle)	Sparse-map path
ddf2b7d	for…in over empty _combined / _extendedAttrs + 2× {} alloc	for…in over map keys only
master (#5972)	O(cols) _copyCellMapsFrom loop every scroll	Same O(cols) loop
This PR	_combined sentinel identity check (O(1)); _extendedAttrs = undefined	Inlined O(cols) flag scan (no per-cell function calls)

Takeaways

• master vs ddf2b7d: fix(buffer): clear stale attrs and combined data in BufferLine #5972 correctness work regressed scroll recycle (~3.9× slower at cols=80) while improving sparse vs ddf2b7d's key-copy approach (~1.8× at cols=80).
• This PR vs master: Faster on all four cases; scroll recycle gains are largest (~9–11×).
• This PR vs ddf2b7d: Faster on all four cases — correct map rebuild without paying master's scroll-recycle regression.

CDP results (timeout 1s yes on demo terminal)

Chromium CDP CPU profile, cols=140, rows=38, scrollback=1000, ~2.76s sample window. Attributed time for copy-path symbols (note: V8 often does not nest copyFrom with _copyCellMapsFrom on the same stack frame, so compare the combined copy-path total).

Symbol	master total ms	master %	PR total ms	PR %
copyFrom	46.70	1.69%	62.57	2.26%
_copySparseMapsFrom	68.36	2.47%	44.33	1.60%
_copyCellMapsFrom	310.02	11.21%	0	0%
Copy-path combined	425.08	15.38%	106.90	3.87%

Copy-path combined time: −75% vs master. _copyCellMapsFrom samples eliminated on the blank-line scroll path.

Design notes (alternatives explored)

• undefined for _combined: Regressed sparse copyFrom in benchmarks; not used for _combined.
• undefined for _extendedAttrs: Adopted (addon-image writes maps directly; shared sentinel caused integration test failures).
• Object.freeze on sentinels: Safe with copy-on-write but no measurable perf gain; not adopted.

Testing

• npm run test-unit -- **/out-esbuild/**/BufferLine.test.js — 51 passing
• npm run test-integration -- --suite=addon-image — 209 passing
• npm run lint-changes

  

[Tyriar]

Testing running yes for a couple of seconds, see the total time % for copyFrom.

Before:

After:

[Tyriar]

Before on ddf2b7d (before the correctness changes in b3de9fa: