A regression is a bug that was not in release N - M and was discovered in release N. You are then free to fix N - M + 1 to N. Like you have a testcase that crashes gfortran on 4.1.0, but did not on 4.0.2.
But then, you'll explain people that they won't have a decent fortran compiler in distros before <insert here a date somewhere in early 2007>.
The fortran patches are always fortran-contained, and I think if the community thinks it worth to have a different development model (until some point in the future, defined in advance) why shouldn't it be so?
FX