M3X-1590: Optimize code for better performance

[ragoolkr]

Ticket

Summary

Performance fixes for the structural table redlining introduced in #8 (v2.0.0). Output is unchanged: all specs pass and results are byte-identical to 2.0.0 on every edited input. The only difference is that identical input is now returned as-is, as in 1.6.0.

2.0.0 was ~3× slower than 1.6.0 on lightly edited and keyed tables, and spent 180 ms (200 rows) to 4.3 s (800 rows) on tables that had not changed at all. With this PR, 2.0.0 is faster than 1.6.0 on every table workload measured.

Changes

File	Change	Why it's better
src/htmldiff.ts	Return the input as-is when before === after	Unchanged sections, the common case in a document redline, go from 180 ms (4.3 s at 800 rows) to 0; same output as 1.6.0
src/core/tokens.ts	Track the last < instead of rescanning the whole atomic token on every >; self-closing check by one character instead of a regex	Large keyed elements (the merged table) no longer get quadratically slower with size; main reason "worst" went 139 → 16 ms
src/tables/RowAligner.ts	Read each row once (words, blank, keys, integer exactKey from the sorted word set)	Per-pair cost drops from regex word-splitting to an int compare or Set lookup; exactKey matches exactly when similarity === 1, so output stays byte-identical
src/tables/SequenceAligner.ts	Optional isExact hook, flat Int32Array DP table, push instead of concat in a loop	Cheap exact pass, one allocation instead of n, less GC; removes quadratic array copying
src/tables/similarity.ts	Split cellSimilarity into signatureWords (parse once) + wordSimilarity (Set-based)	Words are parsed once per cell, not per comparison; public API unchanged
src/tables/TableVersion.ts	rowKeysByRow() builds the owner map once, then reads every row's keys	Keyed tables 150 → 11 ms; no repeated owner lookups per comparison
src/tables/TableMerger.ts	Precompute kept-row group signatures; early return when there's no group; find instead of filter()[0]	Removes a changed × kept recomputation; stops at the first match
src/tables/html.ts	Cache the compiled attribute regex per name	No new RegExp in hot loops
src/tables/helpers.ts	uniqueValues via Array.from(new Set(...))	O(k) instead of O(k²), same first-seen order
.gitignore	Ignore generated htmldiff-cli.d.ts	Keeps build output out of git

Results

Measured with the htmldiff perf bench (drive.sh, best of two round medians; Apple Silicon, Node 25, round spread ≤ 8%).

Case	1.6.0	2.0.0 (#8)	This PR	vs 1.6.0
table 200×6, identical	~0 ms	180 ms	~0 ms	same
table 200×6, 10% cells + row add/del	57.6 ms	190 ms	12.2 ms	4.7× faster
keyed table 200×6, 10% cells	47.7 ms	150 ms	11.3 ms	4.2× faster
table 200×6, 60% cells edited	107 ms	20.7 ms	6.4 ms	17× faster
worst: table 300×8 long cells, 60% edited	828 ms	139 ms	16.0 ms	52× faster
document 20 tables 10×4, edits in each	34.8 ms	36.5 ms	16.2 ms	2× faster
text 506 KB, 5% words	1048 ms	953 ms	955 ms	same core

Scaling, one table with 5% of cells edited: 800 rows took ~2.9 s on 2.0.0 and take 58 ms now (≈2× per doubling instead of ≈4×).

[sirenevenkii]

fantastic!

[sirenevenkii]

is it really testing what it states?

[sirenevenkii]

are those 2 tests? is the name correct?

[sirenevenkii]

we never published it, so it can stay 2.0

[sirenevenkii]

forgot to approve