> 
> Providing an explicit cache bypassing method also sounds worrying to me:
> the cache is supposed to be transparent and semantics-preserving.
> 
> More generally, I think adding new features to an importer shouldn’t
> require modifications in this area, as a matter of separating concerns.
> 
> WDYT?
> 
> Thanks,
> Ludo’.

yes I agree, I removed that in the last version! Which also rebases off
the new topological sort procedure. I'll attach the latest patches here,
in case you missed it.