[Yushu-Li]

Existing OpenEvolve evaluators accept a file path, while AlphaEvolve supplies a candidate object. Add an opt-in adapt_openevolve_evaluator(evaluate, objective_metric=...) helper that writes a single Python candidate to a temporary file and returns the nested score dictionary accepted by the upstream SDK's AlphaEvolveExperiment.

This continues the evaluator-contract work in #468 in the reverse direction. It is an independent slice; #468 remains open.

• Support metrics dictionaries and EvaluationResult, preserving numeric metric names and values.
• Ignore candidate path metadata and clean up temporary files on success, exceptions and interrupts.
• Convert evaluator exceptions, explicit error/timeout markers and non-finite scores to a configurable finite failure score. Generic feedback does not copy exception text or artifact contents.
• Reject malformed, multi-file, non-Python and asynchronous evaluator contracts. Omit artifacts with a warning and document these limits, including the absence of timeout enforcement or sandboxing.

The adapter adds no Google dependency or client/authentication behavior. Export bundles, reverse configuration mapping and managed-service execution remain separate work.

Validation:

• 18 new offline tests, including concurrent calls, callable serialization and round-trip use with the existing inbound adapter.
• Full unit suite: 751 tests passed with a placeholder API key; Black, isort and whitespace checks passed.
• Offline integration against the unmodified Google SDK at 674dd5e558e917320c251c085e47fae33e1bbcff: both unchanged OpenEvolve circle_packing and function_minimization evaluators/seeds passed through the real SDK Experiment and EvaluationWorker, and the outgoing payload passed its submission model validation. A local recording transport replaced the cloud client. An intentionally invalid candidate also produced valid failure scores. No network calls, credential lookups or managed-service run were made.

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