281 caption switcher fix

[maboa]

Summary

Three related caption-editor bugs surfaced while testing the Deepgram flow end-to-end.

1. Stale captions after a new transcript (9e1f849)

populateCaptionEditor short-circuits on a non-null captionCache, which is the right behavior for preserving user caption edits across transcript edits — but
the cache was never invalidated when a fresh transcript was produced. After running Deepgram or Whisper, or importing JSON/SRT/VTT, switching to the caption
page showed captions from the previous transcript.

Fix: invalidate captionCache on the hyperaudioInit event, which all three "new transcript" code paths already dispatch. The localStorage load path manages
its own cache (it loads captions alongside the transcript), so hyperaudioTranscriptLoaded is intentionally not used as a trigger. The Regenerate button is
unchanged.

2. Transcribing from caption mode produced no visible loader and no result (1a4a7c7)

Both getData() (Deepgram) and handleFormSubmission() (Whisper) write the "Transcribing…" loader and the final transcript to #hypertranscript. While the
user is in caption mode, #hypertranscript is detached from the DOM — so the loader never showed and the result never landed.

Fix: programmatically click #transcript-editor-btn at the start of each transcribe flow. The button is disabled whenever the user is already in transcript
view (initial HTML state + set explicitly on switch-back), so the click is a no-op in transcript mode and a clean mode-switch in caption mode.

3. Short-caption stop times were too early (aab08a2)

In caption-modified.js, the short-caption path (segment.chars < maxLineLength) used segment.start + segment.duration for the caption's stop time, where
segment.duration is the sum of word durations. When a segment contained silent gaps between words, the sum was shorter than the wall-clock span, and the
caption text visibly outran its time range — e.g., "So group one is the first one." with a real end at 5.04s was emitted with an Out of 3.28s.

Fix: use the last word's actual end (lastWord.start + lastWord.duration) for the stop time, with fallbacks for missing word durations. The long-segment path
already used per-word end times via lastOutTime and was unaffected.

Likely latent before #277 (which bumped the paragraph-break gap from 0.5s to 4.0s); short paragraphs rarely had large internal silent gaps so sum-of-durations ≈
span.

Test plan

• Transcribe a file → switch to caption view → confirm captions reflect the new transcript.
• Edit a caption in caption view → switch to transcript view → switch back → confirm caption edits persist (cache not over-invalidated).
• Click Regenerate → confirm captions rebuild from current transcript and discard caption edits.
• In caption view, start a new transcription (Deepgram and Whisper) → confirm the view switches to transcript and the "Transcribing…" loader is visible →
  confirm the resulting transcript renders.
• Already in transcript view, start a transcription → confirm behavior unchanged (loader visible, result lands).
• Load a session from localStorage that contains captions → confirm captions are not wiped on next tab switch.
• Transcribe content with notable silent gaps within a paragraph → switch to caption view → confirm each caption's Out time matches the actual last-word end
  (not the truncated sum-of-durations).