One could make an argument and use case for complete flexibility.  That any element could have its own chain of insertion effects.  Maybe even different number of effects up to ... pick your ceiling (3, 10?).  And each effect with its own input to the dry element and own input to some arbitrary effect that came before it in the chain.  And feedback loops.  And bypass.  And ...

Ignoring the DSP for just the raw effect workload (pretending that's all OK and doable) I see a big bottleneck in the routing channels.  All of these audio pipelines between.  Like anything else, we're dealing with finite resources.  It's not all just DSP.

Obviously above I contrived a crazily dense spaghetti bowl that would more than max out what could likely be done.  

Not being an insider, I have no real visibility to the actual limitations.   

Korg does things a bit differently.  They say you can have maximum "x" amount of effects and you can split them between Timbres (is it?) how you see fit.  Then again their system isn't really guaranteed so you also have a CPU "global" pool to deal with.  Changing to this would take serious surgery to the ASICs I would think.

I know 128 elements kind of calls for more effects to take advantage of being able to stuff a single Part with more simultaneous sounds.  A starting point would be to take all of the multi-Part Performances and see what the insert effect story is for each Part then see if there's a technical way to provide the resources to insert effects in a single Part to realize a merge without being different from the Multi Part.  Of course effects are only one dimension of this merge but that would be an easy paper analysis or stretch goal to define the parameters.

It may not be where you're trying to go with this, but one "real world" case that calls for an expansion to insert effect counts and routing.

Main 'end goal' is save/load pre-configured effects.

The main thrust of the question was whether it was just overkill to have an effect architecture at the element level or if there was some (and if some, how much) merit to it.

Some effects are better configured based on an individual source sound/element rather than the part as a whole. For those it means restricting the part in order to let the base element control everything. 

If the effect could be controlled at the element level it would be easier for us to create encapsulated 'components' that could be used more modularly and then reuse those base components when building ensembles.

At the part level everything gets mixed together if too many different element types are involved.

Other effects are more appropriate for a higher level in the architecture. For example an effect that takes the size of the room/hall/venue into account can often be done at the performance level. That is especially so for very large outdoor venues, less so for smaller enclosed spaces.

For a smaller space the actual location of the string section (e.g. near one side of the space) might inform on the configuration of an effect that deals with delays or reverb while a brass section in the middle could use a performance-level configuration.

The more useful angle for us is the modularity - being able to use a piece of functionality as close to where it is actually needed as possible.

I think I'd mostly like to see given 128 elements that we could somehow magically have at least 4 sets of insertion effects that would minimally be fixed the way the current ones are (An->Bn or Bn->An or An||Bn) where n is 1-4 and each element can be assigned to start at any of these 8 insertion effects.  I can't remember if AWM2 has parallel insertion effects or not so scratch the last A||B if parallel is only an FMX feature.

So not unlimited routing, just 4 sets of what we have per Part today.  And this may not be possible but would be kind of a low end of a wish to be able to start stuffing 4 Parts into 1.  Of course there's a lot more to this story outside of insertion effects.