Eval: add retry/iteration support for flakiness tracking

[olaservo]

Problem

LLM evals are inherently non-deterministic. Running each eval once gives a pass/fail but no signal about reliability. A task that passes 2 out of 3 runs is very different from one that passes 3 out of 3.

Currently:

• Single-run only (no `--iterations` flag)
• No pass-rate tracking
• No aggregate statistics across repeated runs

Suggestion

Add optional multi-run support:

• `--iterations=N` flag (default: 1 for backwards compatibility)
• Track pass/fail per iteration
• Report aggregate: `3/3 passed (100%)` or `2/3 passed (67%)`
• Save aggregate results alongside individual run logs

Even N=3 would provide meaningful reliability signal without excessive cost.

Files

• `evals/eval.ts` (parseArgs, main loop)
• `evals/lib/metrics.ts` (aggregate result type)