First of all, I have never seen such a weird issue.
Did you exclude CMAQ spin-up runs which includes significant IC contribution when you calculate average for 3-month CMAQ outputs? Because your ICON source apportionment map shows that O3 (grid concentration/sum of all grids’ concentration) over California has the same magnitudes as your CT-O3 and NY-O3 ( the second and third maps), that implies that your results do include strong ICON memory.
You can try to run 2-month average (July and August), then you may see more reasonable results if you run CMAQ continuously from June 1.
CMAQ outputs will lose ICON memory after15 - 20 days for this case. For long-term CMAQ runs, in my opinion, the impact of ICON on surface O3 can be ignored.
Did you take a look at 3-month average of O3 concentration instead of the concentration divided by domain sum of all grids concentration?
Otherwise you may need to double check your masked region files (CT and NY) used in your emission control file.
I hope it is helpful to fix the issue.