Comment 1 by Stephen Leake, Oct 27, 2010

this test is also broken in the same way in version 0.48.

overriding the execute attr init function lets the test pass in 
8f507884dab9bab6b439cec51f6adc6d006d3e66:

attr_init_functions["mtn:execute"] =
   function(filename)
        return nil
   end

Comment 2 by Thomas Moschny, Oct 27, 2010

How can "attrs can be set after a commit"?

Or, more precisely, how's that different from "file contents 
can be changed after a commit"?

Attributes are part of the manifest and thus of course part of the 
revision hash. Do you propose to change this?

Comment 3 by Richard Levitte, Oct 27, 2010

Looking at the source, I'd say that the problem lies in the added 
lua routine called is_executable().  The actual function that 
implements that routine can be found in unix/process.cc and 
win32/process.cc.  win32/process.cc does nothing (it just returns 
false), while unix/process.cc does what's sensible for Unix.

So, the question in the end is, which of those two does Cygwin use?  
I'm guessing it uses unix/process.cc.  Doesn't seem to be the right 
thing to do, but I don't know enough about Cygwin to have an 
opinion, all I wanted was to point at where the problem should be.

Comment 4 by Richard Levitte, Oct 27, 2010

Further analysis shows that, in fact, Cygwin is regarded as unix by 
configure.ac.  This is probably normally a good thing, since it 
tries to emulate unix...  but when it comes to is_executable, maybe 
that's not the best idea.

So we have a choice.  Either add *-pc-cygwin in the case that 
catches Win32 platforms in configure.ac, or make is_executable in 
unix/process.cc a special case for Cygwin.

Comment 5 by Stephen Leake, Oct 28, 2010

> How can "attrs can be set after a commit"?

'mtn attr set'

> Or, more precisely, how's that different from "file 
contents can be
> changed after a commit"?

I thought it was; apparently I'm wrong.

> Attributes are part of the manifest and thus of course part of 
the
> revision hash. Do you propose to change this?

No, I was just wrong; attrs are part of the revision hash.

This is discussed in monotone.info; I've just missed it before now.

So the issue resolves to ensuring that attrs are consistently 
applied
within a project, and specifically for this and other tests.

There is no way to make an automatic execute attr detector get the
'right' answer on all platforms (because there is no 'right' 
answer), so
that's not a solution for this test. Thus the issue of the execute 
attr
being set incorrectly on Cygwin (and other platforms) is a separate
issue; I'll file a separate bug report for that.

The attr_init_functions I provide below is a reasonable solution for
this bug. It should be in test_hooks.lua, so all the tests
benefit from it.

We should do the same for 
attr_init_functions["mtn:manual_merge"], so
future changes to that don't break tests.

Those are the only attr_init_functions in std_hooks.lua