To disable interrupts when manipulating a "critical" variable is none the worse than the present state, when whole functions are disabled (by explicit use of appropriate pragma in both compilers) when they contain operations with "critical" variables.

Maybe I won't answer all your questions, but I try at least:

1. first step is to implement "critical" as a keyword for variables; using these in "main" would (note, the user assigns the variable as "critical")
   
2. second step is to detect "possibly critical" variables and issue a warning (this can be suppressed using appropriate pragma - there are a lot of such warning-suppressing pragmas in all C-compilers)
   
3. third, a possibly dangerous step is to switch on the automatic "critical" flagging, but this would be by default off and potentially locally suppressable - OK this can be completely omitted if you think it brings more trouble than help
   
4. anyway, I'd implement the "disable" and "enable" as a macro or a function (possibly inlined), so that the user can override it as he wishes

I wouldn't be afraid of "incomplete non-atomicity detection" accusations, if this all is properly documented and if the warning says something like "possibly critical variable" or so. I have seen a lot of such automatic detection and warning on e.g. of "unused" variables, which can be easily cheated, and nobody complains.

I think people are happy if they are helped and I believe even seasoned programmers like to be helped. I might be wrong of course, due to my innate naivity and stupidity :-)

JW