Potential bug: custom feature_dimensions provided via program.metrics should not be binarized

[agadetsky]

LLM prompt optimization example considers using custom features as MAP-Elites coordinates:
https://github.com/codelion/openevolve/blob/599cd54036745a6f4e4d22a74760a77e22778fca/examples/llm_prompt_optimization/evaluator.py#L62

As we can see from the implementation of calculate_prompt_features, function returns bin indexes rather than raw scores.
This will cause some undefined behavior, i.e., further binarization in database.add method:
https://github.com/codelion/openevolve/blob/599cd54036745a6f4e4d22a74760a77e22778fca/openevolve/database.py#L718

From the modeling perspective, I suggest to move binarization entirely out of ProgramDatabase class, since, in practice, really useful feature coordinates are very problem specific and should be defined by the user. Subsequently, user should define how to binarize them and provide bin indices in program.metrics instead of raw scores.

Then, in database._calculate_feature_coords we could just do

elif dim in program.metrics:
    coords.append(program.metrics[dim])

[codelion]

Thanks for reporting, I will fix the example for now and add a note to documentation to remind users on returning the raw scores.