rdmbair15m5-changelog-20260822-0514-local-llm-benchmark-m1-engines

Implemented Milestone 1 engine adapters and telemetry for the Local LLM & VLM Benchmark framework on Apple Silicon.

Scope

pyproject.toml: Package configuration with core dependencies and optional extras (mlx, llamacpp, all).
src/__init__.py: Package initialization.
src/engines/__init__.py: Engine adapter module exports.
src/engines/base.py: Abstract BaseEngineAdapter protocol, BenchmarkResult, StreamingTokenMetrics, HardwareSample, ModelConfig, BenchmarkMetrics, and image conversion helpers.
src/engines/ollama_adapter.py: HTTP adapter for Ollama with streaming timing extraction, structured JSON schema support, and multimodal vision processing.
src/engines/mlx_adapter.py: Apple Silicon MLX adapter using mlx-lm and mlx-vlm, lazy imports, unified memory telemetry via mx.metal.get_peak_memory(), and streaming token metrics.
src/engines/llamacpp_adapter.py: Metal-accelerated llama.cpp adapter supporting llama-server, llama-cli, llama-cpp-python, GBNF JSON grammar constraints, and multimodal projectors (--mmproj).
tests/test_engines.py: Comprehensive test suite verifying all adapters and protocol behaviors (33 unit tests, 100% passing).

uv run pytest tests/test_engines.py -v: 33 passed in 0.15s
uv run pytest tests/tier1_features/test_f1_adapter_protocol.py tests/tier1_features/test_f2_ollama_adapter.py tests/tier1_features/test_f3_mlx_adapter.py tests/tier1_features/test_f4_llamacpp_adapter.py -v: 21 passed