A small update,

> Yes, but I still prefer the delimited format for readability.

Ok: expected output is now between check-output-start / check-output-end
tags.
I've added a clean-check target in the Makefile, updated the provided
test-cases and documentation.

> You can't redirect two things independently to the same file; that will
> open the file twice, and the writes will conflict, giving exactly the
> result you saw.  > o 2>&1 should work; it has exactly the same effect as
> &> .

For some reason I was expecting that:

[pid  8877] open("o", O_WRONLY|O_CREAT|O_TRUNC| \
        O_LARGEFILE, 0666) = 3
[pid  8877] dup2(3, 1)                  = 1
[pid  8877] close(3)                    = 0
[pid  8877] open("o", O_WRONLY|O_CREAT|O_TRUNC| \
        O_LARGEFILE, 0666) = 3
[pid  8877] dup2(3, 2)                  = 2
[pid  8877] close(3)                    = 0

and

[pid 10777] open("o", O_WRONLY|O_CREAT|O_TRUNC| \
        O_LARGEFILE, 0666) = 3
[pid 10777] dup2(3, 1)                  = 1
[pid 10777] close(3)                    = 0
[pid 10777] dup2(1, 2)                  = 2

followed by:

[pid  xxxxx] write(2, "validation/preprocessor19.c:4:9:"..., 73) = 73
[pid  xxxxx] write(2, "validation/preprocessor19.c:3:9:"..., 66) = 66
[pid  xxxxx] write(1, "\ny\n", 3)        = o

to give the same result and that the first one would not eat characters.
I was wrong.

-- 
Damien