It sounds like simple or optimized graph algorithms could do the job, so yeah, why not do it.
What is the obstacle, then? Is it not performant in terms of compuational requirements? Not performant in terms of the quality of the result?
Like I said, any common graph-based algorithm will fail when trying to compute a route based on cycleways alone, so I guess the approach would have to incorporate some sort of gap consideration – and then highlight the cycleway parts of the discovered route. But the gaps can’t be too wide.
Bridge finding (not physical) and trunk finding both look like they do what we want, but maybe not.
Here’s an example of cycleways I would like to exclude from lower zoom levels (highlighted). I don’t see any heuristic to apply to exclude them, or include the others…
