v0.38.0

Added

• Sharded GGUF auto-merge (#1893 criterion 2): apr pull now merges a
  downloaded split-GGUF set (model-NNNNN-of-MMMMM.gguf) into a single
  model.gguf so the existing single-file loader runs the model unchanged —
  no inference-hot-path refactor (which would risk all GGUF inference). Pulled
  sharded models are now runnable end-to-end. merge_gguf_shards is
  type-agnostic (copies tensor data by raw byte range → every ggml quant
  type works), lossless on metadata (preserves arbitrary <arch>.* config
  keys via the new GgufReader::from_file_full keep-all mode), bounded in
  memory (streams to disk, holds ≤ one part at a time), and rejects duplicate
  tensors across parts. Parts are deleted after a successful merge.

Verified

• This release was hardened against a multi-agent adversarial verification
  that found 5 release-blockers before publish — most critically that sourcing
  metadata from the architecture-whitelisted reader silently dropped
  gemma.*/phi3.*/deepseek2.*/falcon.*/etc. config keys, making merged
  models of those (mainstream) architectures unloadable. Contract
  contracts/sharded-gguf-merge-v1.yaml (FT-MERGE-001/004/005/006 + 2 kani
  harnesses), including a cross-parser interop test that loads the merged file
  with realizar's real GGUFModel::from_bytes.