Add a model-in-the-loop eval harness

Protocol tests prove the tool works. They do not prove a model can work out how to use it. That is what the tool description and schema are for, and they regress quietly.

Add `tests/evals/`:

- `run.mjs` - the agent loop
- `scenarios/*.json` - a task prompt, setup steps, and post-conditions
- `grade.mjs` - checks post-conditions with direct REST calls

Rules:

- Build tool definitions from the live `tools/list`, so the eval tests what ships.
- Grade on WordPress state, not on what the model says. No LLM judge.
- Cap each run at 6 turns and 1024 output tokens. Truncate each tool result to about 4KB.
- Cache the system prompt and the tool block.

Default model is `claude-haiku-4-5`. Put the provider behind a small adapter so another model can be swapped in.

Write results to `test-results/evals.json`: pass or fail, turns, tokens and cost.

This issue covers local runs only. CI wiring is #18.

Depends on #15 and #16.

Start by reading the existing protocol tests and the requirements in #15 and #16. Build the local harness under tests/evals/: run.mjs should use live tools/list, scenarios/*.json should define prompts and post-conditions, and grade.mjs should verify WordPress state through REST calls. Done means test-results/evals.json records pass or fail, turns, tokens, and cost for capped local runs.

Written by the indexing model from the issue text.