Whisper.cpp custom tokens break fine-tune inference

This appears to be more a tooling contract mismatch than a training failure: custom-token Whisper workflows are ahead of current whisper.cpp defaults. The Reddit report matches known edge cases where custom vocab sizes behave differently after GGML conversion, and active GitHub work on removing hard-coded token and vocab assumptions appears not fully settled. For now, Transformers-native inference remains safer for custom role tags, and production teams should treat tokenizer changes as a cross-runtime compatibility risk.