Tracking down stuttering (deb-12, LinuxCNC-2.9)

Your theory about overruns  with the mesa card can't be vailid. If it does happen, the card will report an erorr finishing read and stop communicating.
There are pins for tmax (the longest time to read and write that are in timer ticks that can be used to check  (divide by the frequency in latency-histogram to convert to usec.)

There is also a packet errror pin that counts how often it happens

I have done a lot of work with bookworm. You might like to read my how to guide.
docs.google.com/document/d/1jeV_4VKzVmOI.../edit?usp=share_link

A good general approach to trouble-shooting is to do a binary search.

As a first step, to check if the motion issue is in step generation or external
hardware you can plot the following error with halscope

"When I'm jogging I'm getting inconsistent speeds and I can also see the current velocity number in the UI dropping when I hear the stepper stutter. "

If the velocity changes, this is not due to motion hardware but likely the
keyboard repeat jogging issue tommylight mentioned.

RodW: Thanks for that document! I used it during my initial Debian 12 setup and it pointed me towards isolcpus setting and tracking down some keyboard options. It would be great if your document was stickied, I found it references in a few threads and it took me longer to chase it down.

I think there are really two issues here. One is that the steppers don't sound right when running under 2.9 but sounded fine under 2.8 with the same hardware. In 2.9 I can hear a second order inconsistent ticking on long rapids (not jogging) done with G0 commands. I don't think this is missed steps, but I haven't done a test yet (it's in my plans next time I get to the machine). My test plan is to do 100 3" back and forths along a 1/2/3 block and see if I end up at the same place at the end of them.

The other issue is that on long jogs (with keyboard) I get inconsistent speeds. This is likely the keyboard issue, though I've tried to solve it by setting kbpoll to low values (tried 1 and 4) and by turning keyboard repeat to high values. My next thing was going to be setting up a debugger to track key-down and key-up events to see if I'm getting spurious key-up events in the middle of a long hold.

I have a long history (about 15 years) with EMC2 and LinuxCNC and have converted a number of machines with both steppers and servos -- this is the first one that is really vexing me. The same hardware (mesa card, power supply, stepper drivers, and steppers) seemed like it behaved better under 2.8. I might setup dual boot with both environments so that I can do a direct comparison.

Plotting the following error will show if G0 motion perturbations are from the
drives/powersupply//timing or LinuxCNC/RT related issues.

I think you are back learning how to use halscope and follow PCW's advice.
Is the keyboard a wireless one by chance? if so replace it with wired.

Many people have had excellent results with Bookworm.
Many people are using Linuxcnc 2.9 without issues (myself included)
Many people are using the 7i96s without issue including with 2.9 and Bookworm.
Everything points to an external issue but you need to fire up halscope to eliminate the software/mesa side of tihngs

If you are not using the spindle encoder, or PCW can give you the ENC inputs of the 7i76e with a bitfile, look at wiring a MPG direct to the Mesa card to avoid the keyboard,

I'm waiting for the PREEMPT_RT patch to be pushed to the mainline code. I think that missed the 6.4 kernel but might be on the 6.5 kernel. That will be a kernel worth building!