[Yushu-Li]

AlphaEvolve-shaped entrypoints currently have no local experiment lifecycle. Add an opt-in openevolve.compat.alpha_evolve facade that imports the existing supported experiment-config subset and runs a synchronous, single-file evaluator through OpenEvolve's public run_evolution API.

The facade supports experiment creation, seed creation, blocking execution, a best-result-only list_programs envelope, and an async-shaped completion check that never starts a second run. Setup performs validation without executing the evaluator or constructing provider clients. Model endpoints and objectives are explicit; supplied Google connection arguments warn and are ignored. Seed scores are re-evaluated, source bytes are preserved in a private temporary copy, and completed/failed/interrupted instances cannot be restarted.

This deliberately small first SDK slice continues #468; it does not close it. It does not install a top-level alpha_evolve alias, provide SDK model/visualization types, list the entire archive, or emulate remote client/controller methods. Only synchronous evaluation with parallel_evaluation=False is supported. Distinct generation/evaluation budgets are rejected; even equal budgets count OE generation attempts, with seed evaluation and retries additional. An explicit iterations=0 allows seed-only execution. No evaluation statistics or hard cost caps are implied.

Validation:

• 19 new offline tests cover lifecycle/error handling, input validation, budget boundaries, exact-source transport and cleanup, provider setup without an ambient key, best-result envelopes, and real OE seed-only/one-generation execution.
• Full offline suite: 836 tests passed.
• Black/isort checks and Frame scans of all five added Python files passed; Python 3.10 grammar checked.
• Separately exercised unchanged program.py/evaluate.py from Google's circle-packing, TSP and signal-processing examples at SDK revision 34cd919193945c6dc81053f9af594031b54f41a8. Each ran seed-only and with one scripted proposal through OE's actual controller/worker. The real SDK model validated every returned program envelope; source hashes stayed unchanged and guarded socket/HTTP/credential lookup attempts were zero. This used a local harness, not the original upstream entrypoints, a real model or the managed service.

No default OE contracts or dependencies change. The walkthrough includes a runnable seed-only example and the supported subset's limitations.