Eval: add batch runner and aggregate reporting

[olaservo]

Problem

There's no way to run all 7 eval tasks at once. Each must be run individually, and there's no aggregate pass/fail summary across tasks.

Currently available:

• npm run eval:greeting (individual)
• npm run eval:code-style (individual)
• etc.

Missing:

• npm run eval:all — run all tasks, produce a summary table
• Aggregate result file combining all task outcomes

Suggestion

Add a batch runner that:

1. Discovers all task configs from evals/tasks/*.json
2. Runs each sequentially (or with configurable parallelism)
3. Produces a summary table (task, mode, pass/fail, cost, duration)
4. Saves an aggregate result JSON
5. Exits with non-zero code if any task fails

Could be a new evals/eval-all.ts or a --all flag on the existing harness.

Files

• evals/eval.ts
• package.json (new script needed)