I'm really closing the loop at last. Short review:

• IRIS 2025.1 is really quicker than 2022.2 on cross routine calls as well as on local calls,
• IRIS 2025.1 is even more (20-25%) quicker when procedures rather than "classic" routines are being called.

Thanks to everyone who responded to my "puzzle", especially to Jon and Vitaliy!
Have a good day!

New test I've managed to write:
a) is very close to real world example where the "puzzle" originates from,
b) is free from file i/o as its speed can differ on docker version from the Windows one,
c) do some (temporary) globals read/write, while all globals involved are entirely cashed.

Its results on IRIS 2025.1 CE/docker impressed me: in this version it runs 4-5 times faster than in native IRIS 2022.1/Windows and doesn't slow down after intermediate call of a big routine function. This is true even when both routines involved (small ztestLib and big ztestBigMac) were generated as classic (non procedure) functions collections. In the case of procedures results were good in both versions involved into my testing.

If anybody wants to reproduce this new test on native Windows or Linux IRIS 2025.1 or 2024.1 version, I would be happy to upload necessary code and data (zipped GOF file of 40MB, 250 after unpacking).