Bootstrap configuration accepts zero group sampling and fails after metric computation starts

Summary

bootstrapped_gmean_metric() validates n_bootstrap and selection_frac, but does not validate BootstrapConfig.group_keep_frac before using it as DataFrame.sample(frac=...) for the outer bootstrap.

A zero group fraction therefore creates an empty bootstrap sample and lets execution continue into downstream merges/concatenation instead of failing at the configuration boundary with an actionable error. Negative values are delegated to pandas and surface implementation-specific errors. This makes an invalid evaluation configuration fail late and inconsistently.

Expected behavior

The metric should reject non-positive group_keep_frac values before starting any bootstrap work, just as it already rejects invalid n_bootstrap and selection_frac values.

Values greater than 1 can remain valid because sampling uses replace=True and can intentionally draw more bootstrap instances than the source set.

Suggested fix

Add a focused validation:

if bootstrap.group_keep_frac <= 0:
    raise ValueError(...)

and regression coverage for zero and negative fractions while preserving a group_keep_frac > 1 control.

Start at bootstrapped_gmean_metric() and review the existing validation for n_bootstrap and selection_frac alongside BootstrapConfig.group_keep_frac. Add regression coverage for zero and negative fractions, while preserving a group_keep_frac greater than 1 control. Done means invalid non-positive fractions fail at the configuration boundary before bootstrap work begins.