Do you mean that for say pitch and yaw at least CIG are using "mouselook"? That would certainly be consistent with observations on pitch and yaw lacking in inertia, whereas roll, which may have been added later (i.e. not native CryEngine), acts as expected...
I posted in here before about how there was only one single dev working on physics and flight model, and the post was ridiculed by the usual suspects. John is the only guy working on this stuff and has been for 5 years.
If you read his document you'll see that IFCS modifies *control inputs*, and if you've kept up with the dev videos you'll know that the control inputs are now read in a batch process, not every frame. The code has actually been shown on a 'Calling All Devs' video 'Everything flight control'.
As an indicator of how IFCS is networked you only need to look at the Spectrum announcement for the 3.2.0 refactor of IFCS, where Baior posted that the flight model had to be tested in a busy server environment, and asked players to test out manoevures in a busy server and fill out a survey of how it feels. Think about that one...
When CIG optimised IFCS by moving its control updates into a batch thread in 3.0, the controls were laggy and in John's words "sloppy". That had to be 'fixed' and 3.2 is the fixed up version of the batch processed IFCS controls. It was not reverted back, it was bandaided into submission.
So you have an input sensitive system being updated in a batch thread whenever the batch gets a chance to execute. The joystick inputs are being read in a background batch process, modified by IFCS, and translated into ship movement.
And by CIG's admission the whole system is sensitive to networking, so that hints at what else it's doing.
As to the mouse inputs - I don't know if those are read in the batch process or if those are being handled properly in a thread dedicated to control. I suspect from your description that mouse input isn't modified by IFCS and gets merged directly further down the line.
It was all pushed back into a batch to optimise it as it was saturating the CPU, so now it only reads control inputs in a 'lazy' background way. And CIG have talked about all of this in videos and even described what they've been doing with it.
IFCS is ten different brands of 'nope' in a big 'nope' bag.