Local validation note: I tested this against the original long Parakeet failure case on my machine (handy-1774637386.wav, 1163.58s / 18,617,280 samples) using the same chunked batch transcription path/config this PR adds for long history retries.

Result: it completed successfully with parakeet-tdt-0.6b-v3 instead of hitting the previous broadcast-shape inference failure.

On my machine that run finished in ~48.5s end-to-end for the transcription step.

So at least for the failure that prompted my earlier #1166, this looks like it resolves the issue in practice.