run_evolution checkpoint/target_score params; propagate llm.provider (#472); background_blur example

[PetrichorHlacyon]

Summary

Bundles three things for the 0.3.1 release.

1. run_evolution() checkpoint + target_score (original PR)

Plumbs target_score and checkpoint_path through the run_evolution() library API. Both already existed on controller.run() but were not exposed, so library users could not resume from a checkpoint or stop at a target score.

2. Propagate llm.provider to per-model configs

LLMConfig pushes shared settings down to each LLMModelConfig via a shared_config dict — but that dict omitted provider. Every model therefore kept provider=None, and LLMEnsemble._create_model (which routes on provider) sent them all to the OpenAI backend regardless of the config. A config with provider: claude_code crashed with a missing-OPENAI_API_KEY error, making examples/claude_code_quickstart unusable.

>>> c = Config.from_dict({"llm": {"provider": "claude_code", "models": [{"name": "a"}]}})
>>> c.llm.provider
'claude_code'
>>> [m.provider for m in c.llm.models]
[None]          # <-- never propagated

Fixed by adding provider to both shared_config dicts (__post_init__ and rebuild_models). Propagation goes through update_model_params(..., overwrite=False), so an explicit per-model provider still takes precedence.

New tests/test_provider_propagation.py covers propagation to evolution and evaluator models, per-model override precedence, the unchanged None default, survival across rebuild_models(), and end-to-end that LLMEnsemble now actually constructs ClaudeCodeLLM. The propagation assertions fail without the fix.

Supersedes #472 (independently reimplemented).

3. New example: examples/background_blur

Evolving a hot function behind a hard quality gate — the score is a speedup, but fidelity is pass/fail, so a fast-but-wrong candidate scores zero. Deterministic and numpy-only (the video sequence is synthesised from a fixed seed: no webcam, GPU, model or dataset).

Demonstrates the cascade (cheap smoke → quality gate → benchmark, so timing is only spent on candidates that are already correct), artifacts that tell the model why it was rejected, and a (complexity, ssim) MAP-Elites grid.

It also documents two traps, both found the hard way:

• The obvious quality gate is exploitable. Grading mean SSIM and worst-frame SSIM lets a "blur frame 0's background and reuse it forever" candidate score 47x while leaving a visible person-shaped ghost — it damages one region, and the frame average hides it. Grading the worst 16×16 region catches it (0.74 vs >0.97 for legitimate approximations). test_gaming.py encodes the cheats as tests so they must lose.
• Timing-as-fitness needs care. Caching a single baseline measurement lets a transiently loaded machine inflate every later speedup (it reported 82x for a program that is really 62x). Baseline and candidate are now measured interleaved, warmed up, best-of-N.

Bumps version to 0.3.1.

[CLAassistant]

All committers have signed the CLA.