Feedback keys this evaluator emits, when statically known.
Used to render per-evaluator experiment progress and to synthesize error
feedback when the evaluator raises. Best-effort: empty when the keys can't
be determined without running the evaluator. Subclasses that know their
keys should override this. The default also resolves evaluators that wrap
a single named inner evaluator exposed as .evaluator (e.g. the
LangChain string-evaluator integration).
