Hm.  The way conditional variables work is that the "conditionality" of the variable is determined when the variable is assigned (in immediate context), not when the variable is expanded.  When we evaluate the variable assignment in "lib: TARGET ?= two" we don't have the inheritance of the "all" target in scope yet, because that doesn't happen until make actually runs the makefile.  So at the time that we evaluation "conditionality" there is no other value in scope.

Now actually I lied a little bit because even if you set "TARGET = global" as a global setting, which clearly IS in-scope when the target-specific variable is set, the result will still be "two".  That's because conditionality only looks at variables assigned in the same scope.

I don't know if this is really a bug or not.  On the one hand it seems like it would be useful to allow conditional assignments such as you describe, but to do that we'd need to change the way variables work to defer the conditional decision until runtime, which means saving the conditional value.  That clearly doesn't make sense for global variables since they can be decided right then, so it would only be for target/pattern-specific variables.  I'd rather do it all or nothing rather than have global settings work one way and inherited target vars work another.

I'm marking this as an enhancement request, although it's probably a bug that, at least, the docs don't make clear how this works.