Good question! Your PR is correct.

For reference, I think this is what the process is - or should be - at the moment:

1. The user tells lasio with the null_policy keyword argument how and where they want NaN to appear in the end result. This may include telling lasio to ignore the NULL. value from the ~V section and instead retain that value.
2. lasio converts the null_policy to a list of regexps and values to find and convert to NaN when parsing the data section from file to memory. (Note that critically this happens during parsing, not after parsing. This is something that prevents a simple switch-over to using pandas.read_csv as a reader.) The conversion to a list of values occurs at reader.get_substitutions. This function also returns the boolean value version_NULL to indicate if the value indicated by the LAS file itself (NULL.) should also be replaced with NaN. At this stage, we don't know what NULL. value, or even that the LAS file defines it. That's the reason for provisional_null existing. It represents NULL..
3. When parsing values in reader.get_data_section_iterative, the regexps and values are applied and a numpy array is returned which may or may not contain NaNs already. It may also contain values which have not yet been identified as needing conversion to NaNs, because we might not have read the NULL. value yet.
4. If version_NULL is true, the final step is to convert NULL. (provisional_null) to NaN. Your PR fixes this step.

Thanks for cleaning up the broken tests.