* long form chunking with vad
* revert some nonsense + better logging
* refactor: score long-form chunks with the selected VAD backend
* refactor: score long-form chunks in the chunker
* refactor: decide decode strategy in one policy module
* refactor: finish every transcription through one path
* refactor: share one chunked decoder across live and batch
* refactor: load the chunking VAD lazily and cut uncut on its failure
* remove some tests
* inline the VAD path lookup and quiet the padded-length log
9e57a0d792