fix Parakeet split retry for long recordings

[ferologics]

Summary

• restore Parakeet recovery for long recordings by moving chunked retry into a dedicated recovery helper
• retry only the current recoverable ONNX broadcast-shape inference failure
• preserve structured transcription results through recovery and merge chunk outputs at the seam instead of flattening to text immediately
• cover the retry driver with scenario tests for recovery, fail-fast behavior, and bounded retry exhaustion

Why

Long Parakeet recordings can still fail with a recoverable inference error like:

BroadcastIterator::Init ... Attempting to broadcast an axis by a dimension other than 1

That currently bubbles up as a hard failure instead of retrying on smaller chunks.

Validation

• cd src-tauri && cargo check
• cd src-tauri && cargo test
• validated through Handy's normal History re-transcribe flow on ~/Library/Application Support/com.pais.handy/recordings/handy-1774637386.wav

Follow-up notes

• seam merging is still heuristic and would be stronger with word-level timestamps during recovery
• once this stabilizes, the recovery helper should move down into transcribe-rs

[cjpais]

Look at my PR for chunking

[ferologics]

I tested #1173 's chunked Parakeet path against the original failing long recording (handy-1774637386.wav), and it completed successfully.

Given that, I think #1173 is the better direction to land:

• chunking/streaming revisited #1173 prevents the failure by chunking long audio before inference
• fix Parakeet split retry for long recordings #1166 adds reactive split/retry recovery after a Parakeet inference failure

So I'm closing this in favor of #1173 rather than trying to merge two competing fixes for the same issue.

The fallback/recovery idea from this PR may still be worth revisiting later, ideally lower in transcribe-rs, if chunked paths still prove brittle.