Wow, thanks for the detailed answer!
I wish i had an arsenal of silver bullets, or that my old experiences counted for something today, but pipelining theory only really takes into consideration some general concepts, and six years out of the loop are too many to pretend i still know anything about how it’s done today.
The pipelining i learnt, besides a few fundamental rules (f.e. rigorous tracking of assets, naming conventions to be defined and respected at all times, etc.), is intended to be adapted quite pragmatically: one does what works in the specific context.
Indeed, some realities i worked for had live referencing of a car model from the modelling department to lighting and rendering, through rigging, animation and look-dev.
Exceptionally elegant, but then parsing the references would take upwards of three hours (!) on a single core, while the render itself would only take ten minutes.
The parsing was unfortunately required *every time* we pressed “render”.
And it was a fairly common occurrence that because of miscommunication, stuff would change in the file while a sequence was being rendered, with with tragi-comical consequences.
In other places, we had a very firm check-out system, where a department would save out and “publish” as usable a file, and the departments downstream would be able to use it.
Each change would have to be picked up by a third party (script, management system, user.), and passed through the pipeline.
This meant many more files on disk, more work from the users, but also a final LnR scene which would contain the leanest of data, and render/iterate very quickly.
There are infinite shades between the two Maya methods, but if you guys don’t have someone whose time is entirely dedicated to finding out, it’ll be hard to fathom which way is best.
Not because you guys can’t get it, but rather because of the time you *have* to keep investing producing the goods. 
I understand i may have a former invested interest, having been an LnR and Pipeline TD for a living before being hired here, but figuring you guys out would be a task a TD would lap up and run with: it’s what made my former job interesting and worth spending energy for.
Sure, finding the right person, and paying for the right experience they may have accrued, won’t be cheap.
Nor there are guarantees they’ll be a fit, or that they may be indeed able to improve on the issues, or even that you guys will want to follow through with what they’ll suggest.
But temp/trial contracts exist for a purpose, and at the very least, and perhaps at some cost, you’ll have had a bunch of new ideas and avenues to explore.
For the record, i’m a huge fan of Houdini.
But that too is a tool for which you could do worse than to hire a pro to show you the ropes *for a year or two*, before you’ll be all be able to walk on your own.
Even if you guys are hyper smart, the sheer amount of UI, methods, nodes, languages Houdini sports will take time to learn to production standards`, and some serious guidance (meaning: in-house, paid. instead of youtube.) can dramatically shorten it.
I say this because in 25 years of using it, the only times i made inroads, truly, were while *talking* to a pro colleague that was using it.
Tutorials too help, but because of Houdini’s vastness, they are often not of general application, and often aren’t well translatable to other tasks, making the hired professional the most efficient path to productability.