Unfortunately, we must not forget forget ABOMINATIONCase and NAMESPACED_PascalCase (we could also generalize this to any switch from one convention to another after the first word). I've found they usually are, in fact, identifiable and roundtrippable. I've found the following usually works (it fails for multi-word prefixes or non-leading exceptions, and of course single-word identifiers are ambiguous):

  count and strip all leading, then trailing, symbols (in some languages this is not limited to underscore but said languages usually need special handling anyway)
  for every chunk separated by underscore, space, or hyphen (this may be nothing):
    if the chunk either has no uppercase or has no lowercase, simply use it as a word. Otherwise:
    for every sliding-window pair of letters (c, d) in the chunk:
      if c is uppercase:
        if d is lowercase, or d is a digit and there is lowercase elsewhere:
          start a new word before c

  space and kebab case don't have any good answer for affixes AFAIK. Otherwise, restore at least the leading underscores (trailing underscores are usually keyword-avoiding)
  for snake and screaming case, be sure to prepend an underscore if the *result* would start with a digit
  for camel variants, prepend an underscore to each *word* that starts with a digit. But if there were originally more than 1 leading underscores, use those instead for the first word.