Separate generation conditioning from decoder reference context
When a voice-cloning option changes both generation conditioning and waveform-decoder context, improved audio does not identify which change caused it. Hold the generated codec trajectory fixed while changing only decoder context, then separately swap the conditioning example whi…