Add multi-adapter registry with CompositeAdapter and Together AI adapter

Fix #149: MCP protocol server — FastMCP over stdio for agent access

Fix #148: Code review fixes for ADR-013 implementation

Fix #148: Semantic-chunker MCP primitives (geometry, trajectory)

Fix #145: Replace ReactRunner with execute_test_case() dispatch

Fix #144: Extract stateless react_step() MCP primitive from react_execute()

Fix #143: Add ReactRunner orchestration and handler routing

Fix #142: Extend BenchmarkCase with mode, mock_tools, max_iterations

Fix #141: Implement react_execute() MCP tool for ReAct loop evaluation

Fix #138: Cache drift availability check and add sibling repo fallback

Fix #137: Port cycle detection and ReAct schemas from LAS

Fix #136: Surface structured tool calls via __TOOL_CALLS__ sentinel

Fix #135: Support pre-defined multi-turn conversation history in BenchmarkCase

Pipelined judge scheduling (#131) and embedding-based drift validation (#132)

Prevent JIT model swap mid-stream on multi-GPU setups

Allow concurrent requests per server (#129)

Add HuggingFace adapter and Spaces deployment

Document semantic validation and promptfoo handling

Add verdict validation to semantic validator

Fix #123: Map expected_verdict and category from promptfoo vars

Fix #121: Remove queue timeout to prevent cascade failures

Support promptfoo prompt objects (id, label, raw)

Fix #120: Judge dropdown ignores Only Loaded filter

Fix #119: Only Loaded filter returns all models when none loaded

Fix #118: Return inference latency from adapter layer

Fix #110: Update tests for server affinity removal

Fix #110: Remove server affinity prefix system

Fix #106: Update test_server_affinity to use InferenceTask

Add focused GPU1 routing tests

Fix #104: Remove battery semaphore to enable parallel GPU execution

fix: Remove per-request manifest refresh serialization

Fix dispatcher race condition and refactor for type safety

fix: Remove adapter-level lock for parallel multi-GPU dispatch (#104)

feat: Add PromptfooLoader for YAML test files

feat: Wire judge() MCP primitive into BatteryRunner (#102)

test: Server affinity parsing + LMStudioAdapter boundary tests

refactor: Rename Test* classes to avoid pytest collection warnings

feat: Complete multi-GPU parallel execution with server affinity

wip: Multi-GPU parallel execution with server affinity

fix: Battery loop order + adapter race condition

feat: Column-per-GPU model selector

test: ADR-006 test suite alignment - mock at layer boundaries

refactor: ADR-006 Battery tab - MCP registry pattern

refactor: Remove Gemini CLI, update tests to mock MCP layer

chore: Remove obsolete Gemini/FARA adapter code

refactor: Adapter pattern - LMStudioAdapter owns httpx, ComparisonSession uses MCP

refactor(ui): UI Surgery - shared header, remove Stability tab

refactor: BatteryRunner owns pool directly, uses list_models MCP

feat(mcp): Add complete and judge primitives, rename dispatcher

feat(mcp): Add list_models MCP tool - first primitive