Third round of libse micro-optimizations (BenchmarkDotNet-verified)

[niksedk]

Same method as #12530/#12533. Three optimizations, each with A/B output-identity proof and BenchmarkDotNet numbers (.NET 10, Apple M4).

1. Hoist 15 method-local new Regex(...) constructions in subtitle formats — up to ~8,000×

Eight format files constructed regexes inside methods. The worst was WebVTT.RemoveNativeFormatting: three RegexOptions.Compiled regexes built on every call — compiled construction emits IL, costing ~1.7 ms before processing a single cue, on one of the most-converted formats there is. InqScribe built one regex per text line; DvSubtitle six per call; TimedText10/TimedTextNoNs, Son, DCinemaSmpte2010 and the ASSA drawing-tag helper had one or two each. All are now private static readonly ... RegexOptions.Compiled, matching the codebase convention.

Method	Mean	Ratio	Allocated
Old (3 compiled per call, WebVTT shape)	1,738,436 ns	1.000	43,060 B
New (statics)	207 ns	0.000	0 B

2. HtmlUtil.RemoveOpenCloseTags — 1.5× faster, no-op lines allocation-free

Used per paragraph by a dozen format writers (PAC, Cavena890, TS, BelleNuit, several Unknown formats, translate formatting). The Regex.Replace + MatchEvaluator shape allocated a closure over tags, a string per match for the tag-name group, one more per kept match — and rebuilt the string even when nothing matched. Now walks matches, compares the group via span (OrdinalIgnoreCase), and returns the original instance when no tag is removed.

Method	Mean	Ratio	Allocated	Alloc Ratio
Old	1.75 µs	1.00	6.63 KB	1.00
New	1.17 µs	0.67	6.19 KB	0.93

(The alloc delta looks modest because the mixed benchmark is dominated by removal-heavy lines; untouched lines drop from full-rebuild to zero.)

3. StringExtensions.ToRtfPart — 1.6× faster, −52% allocations

sb.Append("\\u" + Convert.ToUInt32(c) + "?") allocated two strings per non-ASCII character — painful on CJK/Cyrillic text in the RTF format/copy paths. Now appends the parts separately.

Method	Mean	Ratio	Allocated	Alloc Ratio
Old	1.90 µs	1.00	15.36 KB	1.00
New	1.21 µs	0.64	7.36 KB	0.48

Correctness

• RemoveOpenCloseTags: 30,000 randomized tag-soup inputs (spaced tags < / i >, unclosed <unclosed, case variants, 5 < 6 non-tags, <v Speaker Name>) across five tag sets — byte-identical vs the old implementation, plus a reference-equality check on the new no-change fast path.
• ToRtfPart: 30,000 random inputs mixing RTF-special chars (\, {, }), CRLF, and CJK/Cyrillic/Greek — byte-identical.
• Regex hoists: pattern strings copied verbatim; behavior covered by the format round-trip tests.
• All 485 libse + 233 seconv tests pass.

🤖 Generated with Claude Code