In a patch build here testing an upgrade of MSVC, we observed some failures in idl_test. From some digging, the problem is that we're taking the wrong branch in the initialization of the _serializationContext member of BypassReplyStruct. The generated code for that looks like this:

_serializationContext(serializationContext.value_or(_isCommandReply ? SerializationContext::stateCommandReply() : SerializationContext::stateDefault())),

_isCommandReply is a static constexpr bool member and is true, but we take the false branch (confirmed through some logging in a local run). This only happens with optimization enabled. If I add another pointless int member to the class and make its initialization dependent on _isCommandReply, then we start taking the correct branch.

These observations may point to the compiler inferring UB somewhere if _isCommandReply is true.