Latest 2021.4 previews now contain the new compiler ideas as discussed further up in this thread. Definition and implementation are decoupled. Most of the modifications done to a patch will only affect its implementation while a few will also affect its definition (like adding a new pin or renaming an operation). Only then dependent patches will need to be looked at as well by the compiler.
Thanks to this strategy the overall workload of the front and back-end compilers decreased drastically while patching. Also the hotswap is much simpler - it no longer needs to traverse the runtime graph, instead it knows exactly what targets of what proxies need to be switched.
The only drawback so far is a loss in runtime performance due to the needed indirection in the emitted target code. Should this become an issue for you please report (our thinking was that this will become less of an issue once we allow to pre-compile VL libraries).
Note that those indirections will not be present in an exported application.