v0.37.0

Added

• Sharded GGUF pull (#1893, pull-side): apr pull now detects and downloads
  COMPLETE split-GGUF model sets (<prefix>-NNNNN-of-MMMMM.gguf). Modern 7B+
  GGUFs ship split with NO index.json (unlike sharded SafeTensors), and apr pull previously ran them through select_best_gguf — silently grabbing a
  single part and producing a broken/incomplete model. Now resolve_hf_model
  detects the complete -of- set (rejecting single-file, multi-quant, and
  incomplete sets) and run_sharded_gguf downloads all parts via a no-index
  path (no SafeTensors conversion), pointing usage at the first part. Contract
  contracts/sharded-gguf-pull-v1.yaml (6 falsifiers FT-SHGGUF-001..006 + 2
  kani harnesses, all passing).
  • Scope: this is the pull side. Cross-shard inference in
    aprender-serve (reading split.count and loading tensors across parts so
    apr run/apr serve work on a split GGUF) is the documented follow-up
    (#1893 criterion 2) — the next release.