With this pulled, test-stamp-propagation started failing even more reliably (on Windows only), usually on the first or second lisp being run. 

I remembered reading somewhere (I think some things Gary Byers was saying on openmcl-devel)  that Windows does not always have particularly fine-grained timing facilities. So I thought maybe 2 seconds sleep was not enough to wait before touching files, and increased that to 5 seconds for Windows, as reflected in the attached patch.

With this change, I ran test-stamp-propagation on all the platforms three times through, all passing. 

For good measure I also bumped the sleep for non-os-windows to 3 seconds from 2. 

Maybe it's not necessary to clear the build/ directory before each set of tests anymore. I'll play with that a little bit. 

 

Oh ya.  It does work when the lisp is invoked from cygwin. So touch is not assumed to be there in normal asdf, only during the tests, right? Then no prob.