nanocodec_decoder.mlmodelc: re-export with compute_units=CPU_AND_NE + FP32 precision (Kokoro ANE recipe). Strictly better than prior export — clean audio on both ANE and GPU; the prior FP16 export had high-frequency hiss when run on ANE.

nanocodec_decoder_ane: re-export with compute_units=CPU_AND_NE + FP32 precision (Kokoro recipe for clean ANE audio)

decoder_step_stateful: KV cache as CoreML state (iOS18+, ANE-resident, no per-step IO transfer)

decoder_step: revert to fp32 outputs (fp16-out caused CoreML to auto-cast on next-step inputs, ~+5-8% RTF regression)

decoder_step: re-export with fp16 outputs (boundary cast removed, halves KV transfer)

Add 8-frame nanocodec for low-latency streaming

Add files using upload-large-folder tool

initial commit