Fix ten more bugs from a second random sweep

[niksedk]

Follow-up to #14025 — a second random sweep, different areas. Nine of the ten were reproduced with a failing test before the fix (regression tests in tests/libse/Core/BugHunt20260823Round2Test.cs, verified red against the pre-fix tree via git stash).

Crashes / data loss

1. DVD Studio Pro "with space" readers threw on a valid file — the regex declares variable-width time codes (\d+) but the parsers cut fixed offsets (Substring(0, 25) / Substring(27)). 0:0:0:0 , 0:0:0:1 , Hi gave ArgumentOutOfRangeException. All four variants now split on the separator, which also stops a separator inside the text from truncating it.
2. Csv silently dropped any line whose text contained ; — it split on every separator and required exactly 4 parts, so a 5-part line was skipped with no _errorCount. Round-tripping "Hello; world" through Csv lost the line. Now split with a limit of 4 (which also stops RemoveEmptyEntries from dropping an empty-text row).
3. ASSA resampling silently gave up on \pos (10,11) — the regexes allow \s* but the offset math didn't, and the failure path was break, abandoning every remaining tag on the line. Parameters are now read between the parentheses, and a bad tag is skipped instead of aborting.
4. WordsPerMinute returned Infinity — no zero-duration guard, unlike GetCharactersPerSecond directly below it. One zero-length line made the average and maximum WPM of a whole Statistics / batch-statistics report ∞ (a reversed line gave a negative rate). Same < 1 guard, same 999 sentinel.

Wrong or dead behaviour

5. WebVTT thumbnails rejected every .jpeg sprite sheet — LoadSubtitle only accepted .png/.jpg.
6. The .jpeg branch of GetBitmap used posJpg (always -1 there) instead of posJpeg — masked by 5, and it would throw on a short name.
7. Netflix "spell out leading number" was dead off Windows — the pattern hardcoded \r\n while paragraph text uses Environment.NewLine. It now captures the digits (removing the offset arithmetic entirely) and accepts either line break.
8. Netflix max-line-length reported the same fix twice for a paragraph with two over-long lines — the fix re-breaks the whole paragraph, so the second record was a duplicate.
9. Video OCR dropped unreadable frames out of their group — the comment said "treat as part of the current group" but continue skipped the EndFrame update, and EndFrame drives the subtitle's end time.
10. ToRawAss had an unreachable duplicate strikeout branch (dead code; the correct branch 30 lines up already handles it). Also fixed alongside: CheckForErrors cut the untrimmed line after testing the trimmed one — same mismatch class as the ActorConverter bug in Fix ten bugs from a random sweep of libse/libuilogic #14025.

Note on test coverage: item 10's dead-branch removal has no behavioural test by definition; the included test instead pins that StrikeOut is written from the style rather than a constant.

Verification

• libse 1445 passed, libuilogic 504 passed, seconv 383 passed, UI builds.
• The 11 new tests were run against the pre-fix tree: 9 failed, confirming each repro.

🤖 Generated with Claude Code