This was tricky.  GNU make tries to avoid the overhead of computing .VARIABLES unless you need it, and also unless it changed from the last time it was computed.  It used the number of variables in the hash table to determine whether the value needed to be recomputed.  This was find when we weren't removing entries on undefine, but when we started to do that now the value is not reliable: if you remove one variable and add one variable (or any equal number of adds and removes) in between expanding .VARIABLES then it doesn't detect that there was a change and doesn't re-compute the .VARIABLES value.

I fixed this in Git by introducing a "change number" for variables which is monotonically increasing, during both adds and removes.

This makefile demonstrates a bug in the expansion of .VARIABLES:

$ cat Makefile.variables
MY_one := 1
MY_two := 2
$(foreach v, $(filter MY_%, $(.VARIABLES)), $(eval undefine $v))
MY_foo := foo
MY_bar := bar
all:; @echo MY Variables = $(filter MY_%, $(.VARIABLES))
$ make -f Makefile.variables
MY Variables = MY_one MY_two