The fact that samba's sendfile causes problems isn't new to you. I propose that 'use sendfile=no' be put in [global] by default.
samba 3.0.x code already sets sendfile off (no matter what smb.conf says) for clients prior to NT4 and when the local tcp sendfile API return ENOSYS and when smb signing is enabled. This should cover all of the obvious reasons why samba should not attempt sendfile. Are there known problems with the tcp sendfile API in particular versions of the kernel itself? I don't know if the ltp or other tests include a simple functional test for sendfile but that sounds like an obvious next step.
More recent versions of Samba have defaulted to "use sendfile = no" unless explicitly overridden in the smb.conf file. Since we don't set it in the default smb.conf, we're set.