Benchmark: tool calls vs DOM actuation on the example app

The WebMCP pitch rests on reliability numbers (fewer errors vs visual scraping). Let's measure it ourselves, honestly, on `examples/todo`:

- Script an agent loop doing N tasks via (a) `executeTool` and (b) a DOM-driving baseline (playwright + a small LLM, or scripted selectors-with-noise)
- Track success rate + latency + tokens
- Publish methodology + results in `docs/benchmark.md` — including where tool calls do *worse*

A credible, self-critical benchmark is worth more than a marketing claim.