Return nonzero status on CLI inference failure

[dissonancehelix]

What changed

• Make predict.main() return 0 after successful inference.
• Return 1 from both handled inference-error paths.
• Propagate that return value when predict.py is executed directly.

Why

The CLI currently catches IOError and general inference exceptions, prints an error, and then completes without a failing process status. Automation can therefore treat a failed transcription as successful.

The existing error output and traceback are preserved; this change only makes the process status match the reported outcome.

Validation

Added deterministic CLI-contract tests with the model and inference call replaced by fakes:

2 passed

The tests verify status 0 after success and status 1 after a handled inference failure. They do not load a model or process audio.