For this sort of things I'm using an intermediate language (guess it's similar to what OP is talking about) with stack semantics: there are instructions like "PUSHINDENT", 'POPINDENT" and "INCINDENT" - with PUSHINDENT inserted after '(' in ifs, function calls, etc. - ensuring that if line breaks before the POP instruction, indentation would reflect the context.

So, the biggest problem is in assigning weights to the break candidate instructions is this stream. This can only be done by a lot of experimenting, I could not find any formal method or a passable heuristic.