Feature: Artifact side channel

[nerdsane]

Why

AlphaEvolve injects “rendered evaluation results – usually a program, the result of executing that program, and the scores assigned by the ­evaluate function” into every prompt, arguing that rich execution feedback speeds convergence.
OpenEvolve presently forwards only the metrics dictionary with numeric values, so LLMs never see build logs, failing-test traces, perf profiles or any other output that may be useful.

In an example I'm currently experimenting with I used this to pass formal model (TLA+) check output and memory profiling results to the LLM. (This is currently work in progress and not included in this PR, so I added the circle packing example with slightly updated evaluator to show how artifacts work).

What’s in this PR

• EvaluationResult dataclass — retains the original metrics dict and adds an optional artifacts field for text / binary payloads.
• Two-tier storage — artifacts ≤ 32 KB are JSON-encoded in a new artifacts_json column; larger blobs are written under artifact_dir/ on disk and referenced from the DB.
• Prompt support — templates now accept {artifacts}; the sampler injects sanitized, size-capped content so the LLM sees exact failure text without blowing context.
• Config & env flags — ENABLE_ARTIFACTS, max_artifact_bytes, and base-path knobs let users toggle or tune the feature with zero code changes.
• Examples & tests — updated circle-packing example shows compile-failure recovery; 26 new tests cover unit, integration, and perf to keep coverage green.
• Backward compatibility — plain metrics dicts are auto-wrapped, so existing tasks run unmodified.
• No impact on selection logic — Best program ranking is still based on pure floats; artifacts are for the LLM only.

How it works

1. Evaluator returns

   return EvaluationResult(
       metrics={"build_ok": 0.0},
       artifacts={"stderr": compile_log}
   )

2. DB stores metrics + artifact blob.

3. Prompt sampler tacks on a block like

   ### Last-run stderr
   ...undefined reference to `foo`...
   

   giving the LLM concrete tokens to fix next round while selection logic still ranks on pure floats.

Impact

• Richer context – faster self-repair of broken candidates, ability to steer the LLM with additional data.
• Zero breaking changes – old evaluators are auto-wrapped; artifacts pipeline is fully disable-able.

[CLAassistant]

All committers have signed the CLA.

[codelion]

This PR addresses #37 as well I believe.

[codelion]

Can you please update from main, we merged a couple of PRs that were already in testing.

[nerdsane]

I resolved the merge conflicts that came with the changes from PRs #47 and #54, and cleaned up a unit test had a Python version-specific assertion, so tests should now pass.

This PR addresses #37 as well I believe.

Yes, any evaluation errors can now be included in the prompt.

[nileshtrivedi]

It might be good to document the convention for artifact names? For eg: build_stdout, build_stderr, run_stdout, run_stderr etc.