[michalharakal]

Against engine 0.62.0, transformers-only and additive. The Android runtime (runtime-iree-android) gains an external-decoder hook for streaming Moonshine, next to 0.63.2's external encoder.

Added

• External decoder hook (moonshine_stream_jni.c, libskainet_moonshine_stream.so both ABIs).
  • A prefill path ext:<library.so>|<model dir> replaces the prefill and step graphs with a dlopened library exporting moonshine_ext_decoder().
  • ABI 1: create, cross, reset, step, destroy.
  • The streaming loop, greedy choice, token budgets, repeat guard and text handling stay in the runtime. Partial transcripts keep their locked prefix.
  • With both sides external, no params archive is needed.
  • Without the prefix nothing changes: the IREE step moved into one helper, behaviour identical.
• moonshine_stream_check: MOONSHINE_DEC replaces the prefill path.
• KDoc on IreeMoonshineStream for the external decoder form.

Checked on an arm32 device, 444 German commands:

• External encoder + decoder on an accelerator outside IREE: 13.5 % WER, transcript 238 ms after the end of the audio (p50), against 1080 ms with the Vulkan decoder; no GPU use.
• Regression: the external-encoder + Vulkan-decoder path gives 444/444 transcripts identical to 0.63.2.

Release steps:

• Commits: version bump to 0.63.3 (gradle.properties, README, docs BOM snippets) and the CHANGELOG entry.
• After merge: tag 0.63.3 on develop.