Optimize waveform render loop and per-line string utility hot paths

[niksedk]

Performance pass over the AudioVisualizer render path, the MainViewModel position timer, and the libse string utilities called per subtitle line. No caching added per request — these are algorithmic fixes (binary search, early exit, single-pass rewrites, hoisted invariants, static compiled regexes).

All changes verified with BenchmarkDotNet (Apple M4, .NET 10). Each old/new pair was asserted to produce identical output before timing.

Waveform / render loop (runs at ~60 fps during playback)

• GetShotChangeIndex: linear scan over all shot changes per frame → binary search on the sorted list: 1013 ns → 7.6 ns (134×) with 2000 shot changes.
• DrawShotChanges: returns immediately when there are no shot changes (the common case — it previously still rebuilt the paragraph position sets every frame); binary-searches the first visible entry and breaks past the right edge instead of computing X positions for every shot change in the movie.
• DrawParagraph: selection check was List.Contains per visible paragraph (O(selection) each — with select-all on a large subtitle that's millions of compares per frame); now a HashSet rebuilt once per render.
• BuildWaveFormFancy: WaveformColor/WaveformSelectedColor/WaveformFancyHighColor StyledProperty reads hoisted out of the per-pixel loop (they went through Avalonia's value store once per pixel column while scrolling/zooming).
• HitTestParagraph: resolves WavePeaks.SampleRate * ZoomFactor once per pointer move instead of twice per paragraph.
• Timeline FormattedText cache gets the same 8000-entry cap as the paragraph text caches (it grew unbounded when scrubbing in frame mode).
• MainViewModel 50 ms position timer: the current-line scan now stops at the first line starting after the playhead (the buffer is sorted by start time) instead of always scanning the entire subtitle.

String utilities (called per line from casing fixes, batch convert, multiple replace, auto-break, grid repaints)

Method	Old	New	Speedup	Alloc
HtmlUtil.RemoveColorTags	1882 ns	282 ns	6.7×	−78%
HtmlUtil.RemoveFontName	2170 ns	358 ns	6.1×	−79%
HtmlUtil.FixUpperTags	548 ns	64 ns	8.6×	−86%
RegexUtils.ReplaceNewLineSafe (\n-only text)	182 ns	73 ns	2.5×	−85%
StringExtensions.CountWords	184 ns	134 ns	1.4×	−82%
StringExtensions.RemoveAndSaveTags (ASSA)	359 ns	198 ns	1.8×	−39%

• RemoveColorTags/RemoveFontName: the regex was re-parsed on every call (uncompiled new Regex(...)) — now static compiled; also dropped the rest-of-string Substring per match.
• FixUpperTags: rescan-from-zero loop (8 IndexOf passes + Remove+Insert full copies per tag) → single forward pass over a char buffer.
• ReplaceNewLineSafe/CountNewLineSafe: skip the two SplitToLines+Join round-trips when the text contains no \r/U+2028 (always true after the first rule in a multiple-replace run).
• RemoveAndSaveTags: input.Substring(index) (rest of string) allocated per </{/\ just for a StartsWith check — now AsSpan. The plain-HTML micro-case is ~13% slower on M4 but allocates 30% less and no longer degrades quadratically with line length; the ASSA case is 1.8× faster.
• CalcFactory.MakeCalculator (hit per grid cell repaint via CPS): dropped the LINQ closure and per-call fallback new CalcAll(); returns the existing shared instances.
• Utilities.CanBreak: dropped the per-call ToArray() copy of the cached no-break list.

Not addressed (fix would require caching, excluded by request): the spectrogram path rebuilding + converting a full bitmap every frame, and the position timer's per-tick copy+sort of all subtitles.

All 1003 tests pass (libse 485, UI 512, libuilogic 6).

🤖 Generated with Claude Code