This is another regression.

I'm going to make a bunch of fixes to runtime.

Once oapi-codegen/runtime#131 is fixed, all the tests here need to be extended to round-trip unicode through all possible parameter encodings and locations.

If we're thorough enough, hopefully we can fix bugs before people hit them.

It's surprising how fragile URL encoding is. I might have to pass some kind of context to the encoder/decoder to know the path something took - how has it been processed so far?