[Yushu-Li]

OpenEvolve experiment settings cannot currently be passed to the AlphaEvolve SDK in the reverse direction. Add an opt-in config_to_alphaevolve helper that exports a documented Python configuration subset as a plain experiment dictionary. This continues #468 after #526; it does not close the full interoperability issue.

• Export the generation-attempt budget and resolved model mixture. Require literal title/problem text, explicit replacement of non-Gemini models, and explicit generation concurrency rather than copying evaluation worker counts.
• Preserve relative weights, merge duplicate target names, enforce the SDK's documented two-model limit, and reject malformed/non-finite inputs. Live model availability and service-specific tier restrictions remain outside offline validation.
• Report omitted local settings by field path, including defaults. Do not copy endpoints, credentials, clients, evaluator controls, MAP-Elites, stopping or prompt-construction settings; do not load templates or mutate the source config.

This adds no dependencies or default behavior changes. It does not export a runnable bundle or create/authenticate/run a cloud experiment.

Validation:

• 19 new offline tests covering conversion, explicit remapping, merging, invalid inputs, omission feedback, credential isolation, no file/network/client activity, and round-trip use with the existing importer.
• Full unit suite: 770 tests passed with a placeholder API key. Black, isort, whitespace and Python 3.10 syntax checks passed.
• Offline verification against the unmodified Google SDK at 674dd5e558e917320c251c085e47fae33e1bbcff: loaded existing circle-packing and function-minimization YAMLs, exported with explicit Python/model choices, validated AlphaEvolveExperimentConfig, and exercised the real SDK Experiment/Client session and experiment-request envelope with only the HTTP transport boundary replaced. Supported fields round-tripped through the importer. No network calls, credential lookups or managed-service execution occurred.

Developed with AI coding assistance; the implementation and validation results were reviewed before submission.