[bpacholek]

Found while translating the captions of a 75 s Polish talk into English on the emulator (Whisper Base, then the offline Polish → English model, #47).

Before: each caption went to Marian alone. Captions often end mid-sentence, so the fragments came back as fragments, capitalised as if each were a sentence:

And here is an explanation, because / Although we will talk about / technical things, / Technological, why her / goes to the tips section, not tech.

After: CaptionSentences.group joins consecutive captions up to a sentence's end (., ?, !, …; at most six captions, so run-on speech stays a reasonable input), and the view model translates those. CaptionSentences.split shares each translation back over its captions in proportion to their length, breaking between words, with a word for every caption while there are enough. TranslateCaptions is unchanged: each copy keeps its caption's time, and its words share it by length.

And here is an explanation, / because although we will talk / about the technical, / technological thing, why does it go / to the tips section and not the tech.

Test plan

• New unit tests in CaptionTranslationTest: grouping on the Polish captions, and splitting by length. :core:model:test, :feature:captions:testDebugUnitTest and spotlessCheck pass.
• On the emulator: the talk's 31 captions translated into English read as continuous sentences. The one stray capital left ("The world,") is where a sentence ran longer than six captions.

🤖 Generated with Claude Code