Release v0.6.2: KV-cache quantization + seq_id branching kwargs

Fix test_latent_primitives for CI without llama-cpp-python

Tests now patch both sys.modules and HAS_LLAMACPP so they work
in CI environments where llama-cpp-python is not installed.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Bump version to 0.6.0, update CHANGELOG

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Release v0.5.1

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Fix transformers 5.4 compat: remove cache_position from generate()

transformers 5.4.0 validates model_kwargs and rejects cache_position
in model.generate() for models whose prepare_inputs_for_generation
doesn't return it (e.g., GPT2). In transformers >=5.0, generate()
manages cache positions internally when past_key_values is provided.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Release v0.4.1

API stability release. 33 issues found and fixed across Easy API, wire
format, connector ABC, and type system. Result objects, CRC32 checksum,
simplified connector ABC. 500 tests pass, cloud validated on A100.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Add CHANGELOG entry for v0.4.0

168 commits since v0.3.2. Major changes: 4 new engine connectors,
3 framework integrations, torch made optional (numpy projection),
deprecated API removed, dead code cleaned, docs rewritten.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Update CHANGELOG for v0.3.2 release, clean up em dashes in docs

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Release v0.3.1: fix protobuf compatibility for Colab/older environments

Remove gencode version check from avp_pb2.py that required protobuf >=6.31.1
at runtime. Now works with protobuf >=4.21 as declared in dependencies.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Mark vLLM integration as experimental with runtime warnings

KV connector plugin (AVPKVConnectorV1Dynamic) has known issues with
PagedAttention format conversion, CUDA graph compatibility, and
concurrent request isolation. Not validated end-to-end with real vLLM.
VLLMConnector text generation works; latent transfer does not.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>