This is an idea.
If we could omit decompilation, it would boost performance.
Therefore, in the long term, we should head towards that state.

One intermediate step could be to rely on the BCEL instead of AST when looking for annotations.
If BCEL won't find an annotation in the class index, it's not in the AST either.

If we could eventually push all of the "locations" searches to pre-decompile rules, we could skip the decompilation for classes without matches.
I mean - we can start somewhere. But maybe it doesn't make sense before solving the harder part.