Remove per-line and per-word allocations from hot text paths

[niksedk]

More of what @ivandrofly has been finding in #13383 / #13301 / #12922: allocations and rescans in helpers that run once per subtitle line, per word or per character. Nothing here changes behaviour — the point is doing the same work with less garbage.

Change casing — the big one

StrippableText's name loop called name.ToLowerInvariant() on every entry of the name list, for every paragraph. For English that list is ~8000 names, so casing a 200-line subtitle allocated ~1.6 million throwaway strings. lower is already lower case, so an OrdinalIgnoreCase search finds exactly the same positions — which is what the loop's own continuation search two lines below already used.

Same method, same loop: the name-end character set concatenated a literal with Environment.NewLine per candidate match, and lower.Substring(start).StartsWith("don't") allocated the rest of the line to test five characters.

Also in this path:

• Two sb.ToString().EndsWith(..) suffix tests inside the per-character casing loop copied the whole accumulated line each time.
• FixCasingAfterTitles took a Substring of the tail for every character position — quadratic on long lines.
• FixCasing ran RemoveHtmlTags twice on the same text per paragraph.

Character counting

CalcCjk.IsCjk allocated a one-character string and ran a compiled regex over it for every character outside two hard-coded ranges — and the grid re-reads CPS and line length on every repaint. It now tests the Unicode block ranges directly.

Because a mistyped range would silently shift CJK character counts, CalcCjkTest.IsCjk_MatchesRegexForEveryChar pins the rewrite to the regex it replaced across all 65536 chars.

CalcNoSpaceCpsOnly / CalcNoSpaceOrPunctuationCpsOnly allocated a fresh calculator on every call, throwing away the memoization CalcFactory goes to some trouble to provide.

Subtitle formats

• SubStationAlpha never got the two fixes [V4+ Styles] already has: style lookup is now a HashSet probe instead of a linear scan of the style list per paragraph, and TrimBuilder replaces sb.ToString().Trim() + newline, which copied the finished output twice more.
• MicroDVD had ten tag branches that each counted a tag over the whole line before the cheap StartsWith that rejects it. Operands swapped, so the common (untagged) case costs one prefix compare instead of ten full scans.
• SAMI: dropped a ToUpperInvariant() copy of each cue that fed an already-OrdinalIgnoreCase search, replaced two Substring(..).ToUpperInvariant() character scans (two to four string allocations per scanned character), hoisted a ~70-char set out of a per-character loop, and built the milliseconds string in a StringBuilder that was already in scope instead of += per digit.
• SubViewer 2.0's IsMine joined the entire file into one string to look for [br], which cannot straddle a line — and .sub is a crowded extension, so this sits on a common probe path.
• Regex.Match(x).Success → IsMatch(x) in 32 places. The first allocates a Match and its group machinery per line to answer a bool.

Hearing impaired / fix common errors / OCR / spell check

• Utilities.IsAllUppercase / HasUppercase replace s == s.ToUpperInvariant() and s != s.ToLowerInvariant() at 11 sites. Both are pinned to the string comparison they replace for every character, plus Greek/ß/accented sample lines.
• The uppercase whitelist HashSet was rebuilt from settings on every line; it is now cached against the list instance it came from, so a settings change still takes effect.
• ReInsertHtmlTags did ContainsKey + indexer per character (auto-break runs this per line, and per keystroke with auto-break-while-typing).
• Helper.FixDash ran RemoveHtmlTags(prev.Text).TrimEnd() twice in one condition and counted at most three lines through LINQ with a TrimStart() string per line.
• OcrFixReplaceList2, all per OCR'd word: four ContainsKey+indexer pairs that each built their key string twice, two inline char[] allocations, and a "\\ell" + postfix concatenation plus full path scan whose answer never changes.
• SpellCheckWordLists built both candidate phrases inside the loop over the user phrase list instead of once.

Benchmarks

BenchmarkDotNet 0.15.8, Apple M4, .NET 10.0.7, default job. The same benchmarks (tests/benchmarks/HotTextPathRound4Benchmarks.cs, added here) run against a git stashed baseline.

Benchmark	Before	After	Ratio	Alloc before	Alloc after
FixCasingNormal (200 lines)	53.05 ms	26.26 ms	0.50	63.16 MB	1.64 MB
LoadSami (500 cues)	1.554 ms	1.211 ms	0.78	4.47 MB	2.04 MB
SubStationAlphaToText (500)	628.8 µs	473.2 µs	0.75	879.7 KB	617.6 KB
MicroDvdToText (500)	189.5 µs	151.4 µs	0.80	338.0 KB	338.0 KB
CalcCjk CountLatin	3.172 µs	2.184 µs	0.69	2496 B	1248 B
CalcCjk CountJapanese	1.251 µs	1.249 µs	1.00	768 B	768 B
RemoveHearingImpaired (500)	1.159 ms	1.128 ms	0.97	2.62 MB	2.61 MB
AutoBreakTagged	10.375 µs	10.251 µs	0.99	11.72 KB	11.72 KB

The bottom three are honestly within noise. Japanese text already hit the two hard-coded ranges in IsCjk, so only non-CJK text gets the win there; the RemoveHearingImpaired and auto-break changes are allocation hygiene rather than a measurable speed-up. Kept because they are strictly less work, not because the table shows anything.

Test plan

• dotnet test tests/libse/LibSETests.csproj — 948 passed
• dotnet test tests/libuilogic/LibUiLogicTests.csproj — 149 passed
• New CalcCjkTest — IsCjk matches the old regex for all 65536 chars
• New UtilitiesCasingProbeTest — both casing probes match the string comparison they replace for all 65536 chars
• Round-trip harness over 14 formats (SSA, MicroDVD, SAMI, SubViewer, SubRip, DVD Studio Pro ×3, DVD Subtitle, LRC ×3, TMPlayer, NVivo), writing + reading back + IsMine, plain and styled input: output is byte-identical before and after (same md5)
• src/ui/UI.csproj builds
• Tools → Change casing → Normal on a real subtitle, and Remove text for HI, to confirm in-app behaviour

Deliberately not in this PR

Avoided Utilities.StartsAndEndsWithTag, StringExtensions, AdvancedSubStationAlpha, SubRip, SubtitleFormat and MergeAndSplitHelper so this doesn't collide with the open #13381 and #13372.

The sweep also turned up a set of algorithmic problems that are too big to bundle here — happy to open them separately:

• TimedText10.MakeParagraph re-runs a whole-document //ttml:region XPath per cue — O(n²) on every TTML save.
• MergeAndSplitHelper.HandleFormatting formats every paragraph from the current index to the end of the file on every translate request (~n²/2 SetTagsAndReturnTrimmed calls per auto-translate run).
• MergeLinesSameTextUtils has no early exit on the time gap, so it is O(n²) — and it runs unconditionally when opening subtitles from MP4/DASH. The UI copy of the same algorithm does break; libse never got it.
• The "Remove text for hearing impaired" window's 500 ms preview timer has no _dirty guard (its sibling FixNetflixErrors does), so it re-detects the language and re-parses names.xml twice a second for as long as the window is open.
• AssaStylesViewModel.UpdateUsages is O(styles × paragraphs) with two TrimStart allocations per pair.

🤖 Generated with Claude Code