Size of JRE.xcframework exceeds 75MB limit of appstoreconnect

I'm using the JRE.xcframework in my own framework. I've built the JRE.xcframework for arm64_32 and armv7k. I can see that that all worked well, because there is a corresponding folder. So my full set up is: My project -> links my framework -> links JRE.xcframework

When uploading my app, is not exceeds the 75MB treshold by a tiny margin:
"ITMS-90389: Size limit exceeded - The size of the watchOS variant at "/Payload/Surfr..app/Watch/Surfr..app" is 75MB, which exceeds the maximum allowable size of 75MB after app thinning. For details about reducing your app's size, view: https://developer.apple.com/documentation/xcode/reducing-your-app-s-size"

I followed the instructions and indeed found out that in some of the thinned IPA's there is a Surfr..app in the Payload/Watch folder of 75.8 mb. This is mainly because of the linked binary of the JRE.xcramework, which is already 68MB (approx 32mb per architecture). If I remove one of the architectures (and build JRE.xcframework, my framework , and my app again) than the size of the Payload/Watch/Surfr..app drops to 43MB.

It seems like my thinned IPA's aren't really thinned/splitted per architecture. I would expect that one of my thinned IPA's would never contain more than one architecture.

What is expected here and how can I resolve this issue so that in my thinned IPA's only one architecture is included and I'm not exceeding the appstore limit?

Here a picture of the linked JRE.xcframework, which make my watch IPA increase with 68MB, because the JRE.xcframework contains two frameworks.

Here the actual problem, a 75.8MB watch IPA within my app archive bundle (thinned ipa's). If I built the JRE.xcframework for, for example only ARM64_32, then the size drops to +- 40MB. But appstore requires both ARM64_32 and ARMv7k so dropping an architecture is not an option. My main question is though, why does including an extra architecture lead to lineair increase in IPA size. I would expect that the app splitting/thinning proces would take care of that and the IPA would be splitted per architecture.