Experiment: does TOON degrade MCP task performance?

[jirispilka]

Problem

TOON cuts bytes, but LLMs are trained on JSON. Switching text to TOON could lower task-success rates. This is the go/no-go gate for #913's TOON PR.

Proposed solution

• Run the existing workflow eval harness (pnpm run evals:workflow, evals/workflows/results.json) on master vs a TOON branch.
• Gate: no task-success regression. Record byte savings on the same runs.
• Capture results for the blog post.

Part of #941. Blocks #913 PR 2.