Can whisper_full reuse the encoder output from language auto-detection for the first chunk? #3954

whisper_full_with_state() first calls whisper_lang_auto_detect_with_state() when the language is auto. That call encodes the first audio window. The main transcription loop then calls whisper_encode_internal() again at seek == 0.

For short dictation clips, this second encoder pass appears to account for almost all of the auto-detection overhead.

I tested a small prototype that reuses the encoder state from language detection for the first chunk, restricted to offset_ms == 0 and the default audio_ctx.

Test configuration: Apple M5 Pro, Metal, Flash Attention, large-v3-turbo-q5_0, five iterations per case.

After reuse, auto detection was effectively as fast as an explicit language on the short clips (265 ms versus 263 ms). The transcript text was byte-identical before and after for all three clips, and the existing transcription tests passed with both an explicit language and auto detection.

This is only a narrow experiment. I have not established that reuse is safe with nonzero offsets, custom audio_ctx, VAD, parallel processing, Core ML, or every callback/backend combination. In particular, reuse would need to preserve the encoder_begin_callback and abort semantics discussed in #3888 and #3936.

Is the second encoder pass intentional because language detection leaves some state unsuitable for decoding, or would reusing its encoder output for the first chunk be valid under appropriately restricted conditions?