Record better circle packing result (2.635977) from issue #156

[akushonkamen]

Fixes #156

What

Lands the circle packing result reported by @yyh-sjtu in #156 into the example's public record, replacing the previous best (2.634292402141039):

• examples/circle_packing/best_program.py: replaced with the program from issue Seems a new circle packing result (2.635977) when reproducing your example 🎉 #156 (verbatim, only reformatted with the repo's black/isort config). run_packing()/visualize() interfaces and the EVOLVE-BLOCK-START/END markers are preserved. A provenance header above the evolve block documents the source issue, the run configuration (phase_1 x100 -> phase_2 x100 -> phase_2 x250, evaluator timeout 200s; best found at iteration 206, generation 13), and the verification command. The program is deterministic — it contains no random initialization (the previous best used an unseeded np.random fallback).
• examples/circle_packing/best_program_info.json: metrics updated to locally re-verified values, with the reporter's run metadata (id, generation 13, iteration 206, timestamps) and a provenance field (the file is only ever written by the library, never read back, so the extra key is inert).
• examples/circle_packing/README.md: ## Results section updated to 2.635977394746595 / ratio 1.0003709277975694 (~0.037% above AlphaEvolve's 2.635); the "matching ... to within 0.04%" narrative updated to "exceeds"; added a reproducibility note stating the program is deterministic and can be re-verified independently with the repository evaluator.
• tests/test_circle_packing_best_program.py: new regression test (unittest, no LLM, no API calls) that loads examples/circle_packing/evaluator.py by path and asserts the recorded program still achieves validity == 1.0, shapes (26, 2)/(26,), and sum_radii >= 2.6359 (lower bound with headroom for ~1e-11 scipy/BLAS float variation). Skipped automatically if scipy is not installed (scipy comes from examples/circle_packing/requirements.txt, not the main dev deps). Adds ~45s to the suite (the program is evaluated twice).

Verification (repo evaluator, no LLM)

$ python -c "... ev.evaluate('examples/circle_packing/best_program.py') ..."
Evaluation: valid=True, sum_radii=2.635977, target=2.635, ratio=1.000371, time=22.72s
{"sum_radii": 2.635977394746595, "target_ratio": 1.0003709277975694, "validity": 1.0, "eval_time": 22.72..., "combined_score": 1.0003709277975694}

This matches the issue's reported 2.6359773947566274 to within ~1e-11 (scipy/BLAS float variation). Validation uses the repository's own evaluator which, as codelion noted on #156, follows AlphaEvolve's released code (google-deepmind/alphaevolve_results B.12) — results validated with it are therefore valid.

Red -> green evidence for the new test

• Against the old best_program.py (unseeded random placement): both test methods fail, e.g. AssertionError: 2.6289329709819897 not greater than or equal to 2.6359 and 2.6265425556110142 not greater than or equal to 2.6359 — the old program doesn't even reliably reproduce its own recorded 2.6343.
• Against the new program: Ran 2 tests ... OK (44.3s), sum_radii = 2.635977394746595 on both runs (deterministic).

Full suite: OPENAI_API_KEY=test-key-for-unit-tests python -m unittest discover tests -> Ran 576 tests in 78.293s ... OK (includes the 2 new tests).

Not included

• The 2.635983 result mentioned later in Seems a new circle packing result (2.635977) when reproducing your example 🎉 #156 by other users: claimed without publishing code, so it cannot be verified or landed.
• No changes to the openevolve core library; no re-running of evolution.

[CLAassistant]

All committers have signed the CLA.