The problem is not to tell selector from definition (of course it can, otherwise grammar would be non-reducible). The problem is how far in an array of tokens the distinction can be seen.

Parsers know the current state and where to go next out of previous reductions and a few lookaheads which they can use at places known to be clear of ambiguity in a ~fixed number of lookaheads. Parsing something that diverges only down few pages of tokens is problematic size/speed/algo-wise, but the question is how much does that cost really and who should pay.

Edit: But honestly I think this is all wrong direction. Any technology, format, method has 1. features and 2. form. Form is irrelevant because you can make tools to convert any form into any. Features are important because you can’t add features by tooling. But $subj adds zero features, only a new form which is already feels like a compromise. Why do that then? We have to brush our tools instead, e.g. create default nginx-scss plugin, standalone scss caching servers etc. Then everyone could use extended syntax without resorting to all-in-one monstrosities like webpack.