Fix 20 bugs found in a bug hunt

[niksedk]

Sweep 18. The verified backlog from earlier sweeps' agent reports, plus fresh passes over the Tools dialogs and the remaining libse areas.

Suites: libse 1635 / libuilogic 708 / seconv 422 / UI 4159.

Wrong results

• Apply minimum gap could write a negative duration. Pulling a line's end back to next.Start - gap can land before that line's own start when the line is short and the next one starts soon after. It was applied anyway and counted as a fix, so the bad duration reached the grid and the saved file.
• Adjust durations (Fixed and Percent) capped the end flat at the next line's start, so two rows sharing a start time produced a zero-duration line and out-of-order rows a negative one. The Seconds branch in the same file already floors its result, and libse's SetFixedDuration / AdjustDisplayTimeUsingPercent leave the minimum gap — so the same operation behaved differently in Batch convert.
• Sort by: removing the last criterion left the previous sorted order in the preview, and OK applied it. ApplySorting already returns the rows unsorted for an empty criteria list; an early return in UpdatePreview stopped it ever running. ClearSortCriteria repopulates by hand for exactly this reason.
• Convert actors applied whatever the 500 ms preview timer last computed, so pressing OK right after changing the from/to type, colour or casing applied the old settings while SaveSettings persisted the new ones — and before the first tick it applied nothing at all. Five sibling dialogs in that folder already refresh on OK and carry the comment.
• Merge lines with same time codes produced a double dash. The default dialog style tested StartsWith("- ") for an existing dash while its sibling case tests "-", so "-Hello" came back as "- -Hello". Fixed in both the libse helper and the live copy in the dialog.
• FirstOnOrAfter/FirstOnOrBefore ran their tolerance probe even on an exact BinarySearch hit and returned the neighbour on the wrong side. A cue already snapped onto a shot change asked for "the next shot change" and got one 20 ms earlier. The generic overloads in the same file only rewrite the index when the search missed.
• StrippableText took the text before [ as [0, idx - 1), dropping the very character the test inspects: "Yes.[Motor roaring] hello." saw "Yes" instead of "Yes." and left hello lowercase. The documented "I [Motor roaring]" case is unaffected.
• The word-split exception tested a lowercase 'a' that the enclosing "starts with its own uppercase" guard had already excluded, so "Acat" was never split while the 'I' sibling ("Iam") worked.
• Google Lens always sent LocaleContext.Language = "en". The chosen language reaches the Accept-Language header but not the protobuf request, which read a config that nothing ever populates (new Lens() takes no config and UpdateOptions is never called).
• The grid's ASSA highlighter treated \clip as a colour tag — the test excluded a digit after c but not a letter — so a {\clip(...)} block cleared the colour state and the rest of the line lost its colour in the grid while libass and SE's own preview kept it.
• Bridge gaps wrote fractional milliseconds; now rounded at the producer (Duration can become inconsistent by 1 ms while editing #14056).
• On Windows, an association the user reassigned in Explorer still reported SE as the default: a non-matching UserChoice fell through to the legacy Software\Classes key that SE itself had written.

Never displayed, never persisted

• 15 buttons set an icon and then bound ContentProperty over it, so the icon never rendered and — since the label starts empty — the buttons had no accessible name either. One did it in the opposite order and lost the bound text instead. UiUtil.WithIconLeftBindText exists for this and is used correctly by five sibling dialogs.
• The three download dialogs set Error on every failure path but bound only ProgressText, so the user saw a generic "Download failed" while the real cause went to a property no window rendered. Three other download windows bind it.
• The auto-cast speakers dialog restored a window position that nothing ever saved.
• Accepting the IndexTTS 2.5 licence was not persisted immediately, so it could be asked again after an abnormal exit. The voice-cloning consent dialog saves on accept.
• The speech-history dialog's play buttons opened disabled until an unrelated watchdog tick enabled them.
• Chatterbox's install-folder path was added one grid row below its own label.

Leaks

Cancelling an engine download stranded the partial .7z — up to ~1.5 GB for Faster-Whisper-XXL / WhisperX — in the engine folder, since only the success path deleted it.

A fix the tests corrected

My first ApplyMinGap patch used the same newDuration > SubtitleMinimumDisplayMilliseconds guard the batch converter uses. That broke two existing tests — correctly: the default minimum is 1000 ms, so the guard skipped ordinary shortening and defeated the feature. Those tests encode the dialog's real contract. The shipped fix guards only against a non-positive duration, which is the actual defect, and a new test pins it (it fails without the fix).

Flagged but not fixed

Two CEA-708 / MacCaption findings need someone who can validate against real .mcc files from other tools, so I have left them alone rather than change what SE writes on a static-analysis argument:

• Smpte291M never skips the optional CDP time_code_section (id 0x71) when CaptionDistributionPacketTimeCodeAdded is set — the reader hardcodes the cc-data offset, while the sibling ServiceInfoAdded flag is honoured. Files from other tools would decode to nothing.
• The SMPTE 291M ancillary checksum is emitted as the literal 0xbb. The arithmetic suggests it should be 0xB4, and 0xBB happens to equal 0x61 + 0x01 + 0x59 where 0x59 is a stale data count the writer later overwrites with 82 — so it looks like a genuine leftover. SE re-reads its own files because the reader never verifies the checksum.

The mpv event-loop test still fails intermittently here; per the previous PR that is a worktree artifact, not a source issue, and nothing in this PR touches the video players.

🤖 Generated with Claude Code