fix: MCP loader/agent — empty-list semantics, collision warning, kwargs forwarding, concurrent execution

[wilsonhj]

Fixes

Four defects in the MCP tool layer, plus a performance improvement:

• MCPToolLoader.load_tools([]) treated an explicit empty list as "no filter" (load everything) instead of "no tools".
• Same-name tool collisions across servers silently dropped tools; now logs a warning naming the collision.
• MCPToolAgent.ainvoke violated its own -> dict return type when no tool was called.
• ainvoke dropped **kwargs instead of forwarding them to llm.ainvoke.
• Independent tool calls now run concurrently via asyncio.gather (order-preserving; per-tool error isolation unchanged — one failing/slow tool cannot affect siblings).

Verification

• tests/unit_tests/tools/test_mcp.py: 11 passed, including a rendezvous test (asyncio.Event-based, no wall-clock sleeps) that only passes under true concurrency — it fails on revert to the sequential loop.
• Cross-checked: the MCP modules depend on nothing outside this diff (loader uses config.data.get(...) dict access; model() call signature unchanged).

Sequencing

Includes the shared dependency pin (needed to build standalone); that hunk rebases away once #10 merges. Otherwise independent of the other PRs in the series.

Co-authored-by: Claude noreply@anthropic.com

[wilsonhj]

Runtime verification — PASS ✅ (real stdio MCP server end-to-end)

Verified against a live MCP server over stdio transport (FastMCP, mcp 1.28.1: add, async slow_echo with 0.5s sleep, fail_tool that raises), real config.yml mcp.servers section, real MCPToolLoader + langchain-mcp-adapters 0.1.14. Only the LLM layer was a fake chat model emitting preset tool_calls; loader, transport, and tool execution were all live. No test-suite rerun.

• load_tools() over stdio → discovered ['add','fail_tool','slow_echo'] ✅
• load_tools([]) → 0 tools (the fix — old code treated [] as falsy = load everything); control load_tools(['probe']) → 3 ✅
• Concurrency, wall-clocked: sequential baseline 1.594s vs agent ainvoke 0.807s — 1.97× speedup (near-perfect 2×), order preserved ✅
• Failure isolation: [add, fail_tool] batch → add result present, fail_tool mapped to an error entry, no exception propagated ✅
• No-tool-call path: plain response still returns a dict ({'content': ..., 'tool_results': []}) — the return-type fix ✅
• kwargs forwarding: ainvoke(..., stop=['END'], foo='bar') arrived at the model exactly — previously dropped ✅

One sub-claim not driven live: the cross-server same-name collision warning (requires two servers sharing a tool name; the logic is a simple dict-overwrite-with-warning). Everything else runtime-confirmed.