Python commands lag time for remap

Hello,
I am converting a Biesse Rover 13s over to LinxCNC and am testing the remap functions with Python and the commands take about 5 seconds to process each one. It's nearly instantaneous from MDI and I don't see any errors in the terminal output when the python command runs. It also works just fine when I reference a straight ,ngc file with REMAP=M201 modalgroup=10 ngc=m201.

Any ideas as to why Python is so slow? I can use the direct ngc remaps but would like to experiment with python also.

Terminal output when running Python
is_callable(oword.remap_m201) = FALSE
Executing remap_m201: Loading m201...
Wait 10 seconds
m201 executed successfully.
task: main loop took 65.045437 seconds

toplevel.py
import sys
import remap

My INI file
[RS274NGC]
PARAMETER_FILE = linuxcnc.var
USER_M_PATH = ngc
SUBROUTINE_PATH = ngc
#RS274NGC_STARTUP_CODE = G21 G40 G90 G94 G97 G64 P0.025
#REMAP=M201 modalgroup=10 ngc=m201
REMAP=M201 modalgroup=10 python=remap_m201
#ENABLE_EMBEDDED_PYTHON=1
DEBUG = 5

[PYTHON]
TOPLEVEL=/home/wrhammer/linuxcnc/configs/Rover13s/python/toplevel.py
#PATH_PREPEND=/home/wrhammer/linuxcnc/configs/Rover13s                                   
PATH_APPEND=/home/wrhammer/linuxcnc/configs/Rover13s/python/     
LOG_LEVEL=10


Python Test Scipt:
def remap_m201(prolog, params=None):
    """
    Python-based remap for M201 to manually load and execute m201.ngc.
    """
    print("Executing remap_m201: Loading m201...")

    import linuxcnc
    cmd = linuxcnc.command()

    try:
        cmd.mdi("M64 P0")
        cmd.wait_complete()
        cmd.mdi("M64 P1")
        cmd.wait_complete()
        cmd.mdi("M64 P2")
        cmd.wait_complete()
        print("Wait 10 seconds")
        cmd.mdi("G4 P10")
        cmd.wait_complete()    
        cmd.mdi("M65 P0")
        cmd.wait_complete()  
        cmd.mdi("M65 P1")
        cmd.wait_complete()  
        cmd.mdi("M65 P2")
        print("m201 executed successfully.")

    except Exception as e:
        print(f"Error in remap_m201: {e}")

    return

The default timeout for wait_complete() is 5 seconds.
Do you need wait_complete?

I don't need wait_complete but was trying different things. With a simple M64 P0 it takes 5 seconds for it to command the pin. Feels like a config somewhere in the python plugin maybe?

remap.py: