Add an mcp level to the agent benchmark

[e54-bot]

Problem

There is no measurement of whether native Wright tools improve on shell + JSON for task success and search/read/tool-call efficiency. See ADR-0020 (Proposed in the ADR PR for #470).

Scope

Add mcp as a wright level beside bin in the agent benchmark condition grid (<wright>/<knowledge>/<network>), with scenario, prompt, grader, model, and trial identity unchanged:

• the harness describes the level to the adapter, and the reference adapter registers the wright serve --transport mcp server for the agent;
• the Wright trace records MCP tool calls so wrightUse, friction, and the report cover them;
• the report compares mcp with bin on usable/passed rate with Wilson intervals and on search/read/tool-call counts, turns, and tokens; tool-schema context counts toward the mcp cost;
• docs/agent-benchmark.md documents the level.

Non-goals

• The docs/skill/Wright+skill conditions and paired-lift reporting (Establish a Wright effectiveness benchmark against docs and skills #466).
• Model ranking or a composite score.
• Changing graders or scenarios.

Acceptance criteria

• agent_bench.py validate and the existing benchmark unit tests pass unchanged.
• A run with level mcp produces result.json with MCP calls counted in wrightUse, and a canary fails the run as invalid when wright is reachable under level none, as today.
• report shows mcp versus bin paired on the same scenario, agent, and trial, with the metrics listed above.
• Tool-schema tokens are included in mcp input-token accounting, verified on a fixture usage log.

Verification

Cases in benchmarks/agent/test_agent_bench.py; one real run recorded in the PR as evidence that the level works end to end (not as a product claim).

Dependencies / ownership

Owner: Wright (benchmarks/agent). Parent: #470. Depends on the MCP adapter Issue. Orthogonal to #466.