Bugfix release (Termux edition, com.termux).
Fixed
- Rate-limited API responses (HTTP 429) now include
Retry-AfterandRateLimit-*headers so OpenAI/Ollama clients can back off correctly instead of guessing. - Attempting to load an embedding model into the generation runtime now returns a clear error; embedding models are served on demand through the embeddings endpoints and no longer need to be loaded.
Debug APKs (5 ABIs) + sha256sums attached by CI.