I see the issue, the main problem being that for no match given the actual capitalization of the input, the closest match for all-lower input is given (completely ignoring capitalization).

The most consistent way around this would be to return the first match obtained with as little capitalization changes as necessary (potentially preferring changing from lower to upper over upper to lower).
However this would mean that 2^(len(units)) changes are evaluated in the worst case. However this may not be a problem as unit strings do not contain more than 4 seperate units

Why not leave capitalization intact for matching a no-match object? Especially for prefixes this is a huge difference.

What do you mean by "leave capitalization intact"? If units were compared without changing capitalization, "mw" would neither match "MW" nor "mW" which are the only valid official versions of writing Mega- or milliwatt respectively.

Switching to all-lower, all of "mw" is matched to "MW" as well as "mW" which both are valid units. I see that another option would be to give preference to the match that is closest to the original string (which would return mW here, as expected)

Then I don't know exactly how quantulum works. I thought there was machine learning doing the matches, not exact matching?

I also learned you will get random output. There is a 50 / 50 one you get either one.

It depends. Most of the tool is a very complex regex-based mechanism, it will turn up with "1pw" or "1mw". The unit in the back is then compared to all known units. If multiple units match, either a simple heuristic or the machine learning part is used to decide which unit has the highest likelihood of being correct.

Hmm random output sounds bad. On the exact same input? Once trained, the output of the trained network should be deterministic.

Yes, random output is bad. At least a warning should be thrown if picking a random option. Anyway, I made a PR.