Explain the cuBLAS compute type failure and offer the fix

[niksedk]

Fixes #13902

The problem

On some GPU/driver combinations Purfview's Faster-Whisper-XXL dies inside encode():

File "faster_whisper\transcribe.py", line 1719, in encode
RuntimeError: cuBLAS failed with status CUBLAS_STATUS_NOT_SUPPORTED

The run produces no segments at all. Subtitle Edit had no probe for that line, so it fell through to "Loading result from STDOUT", found nothing, and closed with an empty subtitle and no explanation — the reporter had figured out on their own that adding --compute_type float16 fixes it, and had been typing it in by hand.

The change

OutputHandler now flags CUBLAS_STATUS_NOT_SUPPORTED the same way it already flags the CUDA out-of-memory line, and the completion handler reports it instead of silently producing nothing.

Since the cure is always the same, the message offers to apply it rather than only naming it: on Yes it appends --compute_type float16 to the parameters box and saves, so the next run works and the setting sticks.

Three shapes of message, so the advice is never wrong:

• faster-whisper based engine, no --compute_type set → Yes/No offer to add --compute_type float16
• --compute_type already present → suggest another value (float16, int8, float32)
• any other engine → no --compute_type suggestion (it would just become "unrecognized argument"); suggests another model/engine or CPU

SupportsComputeTypeParameter gates the offer to Purfview XXL, CTranslate2, and WhisperX — the faster-whisper/CTranslate2 based engines, which are also the only ones that can raise this error.

Testing

dotnet build src/ui/UI.csproj — clean, 0 warnings. The failure needs the reporter's specific GPU to reproduce, so the detection path itself is not exercised here; it is a plain string probe next to the existing CUDA failed with error out of memory one.

🤖 Generated with Claude Code