The download recipe relies on buildout's download, which uses hardlinking for
performance reasons in 2 cases: downloading from cache or from local file.

This means 2 things:
- the user shall not modify downloaded files without first making sure
  that st_nlink == 1
- the source file shall not be modified in-place, which is reasable to assume
  for both cache and local file

In-place modification of files is more and more rare because it has many
drawbacks and it tends to be limited to things like logs and databases.
This commit adds a test to check that Git does not do that when managing
working copies: and actually, this may be our only use case of local download.

This commit does not mean that the current way of hardlinking is fully fine.
There remains at least the issue that file permissions may be changed at the
end of the recipe, either with `shared=true` or with `mode` option.

By default, recent SSH (since v8.6) refuses to connect to not-so-old
server (e.g. the version on Debian 8) with RSA key.

It was partially broken since commit 8a9e3766:
it only works when combined with `environment` option.

These files are usually so small that it's free (< FS block size)
and it is more readable with non-str option values (possible with
slapos.buildout).

This fixes a regression that was introduced in
commit d5f5d399.

Traceback (most recent call last):
  ...
  File "slapos/recipe/build/__init__.py", line 191, in pipeCommand
    raise zc.buildout.UserError('Failed while running command:'
  File "slapos/recipe/build/__init__.py", line 192, in <genexpr>
    + ''.join('\n  ' + q for q in failed))
TypeError: cannot concatenate 'str' and 'tuple' objects

With only 384M, debian-11.0.0-amd64-netinst.iso fails to initialize
the network interface.

Since we switched to running tests with python -m unittest discover,
doctests were not longer ran.

When using init or install option of slapos.recipe.build to
write python code, the traceback were hard to understand, because
the current line was not included in the traceback.

This put the script in linecache before compiling it, so that in
case of error the problematic line appears in traceback.

There is still the "problem" that the traceback line numbers starts
from the beginning of the script and not the actual line from
buildout profile, but this does not seem feasible to adjust the line
numbers, so let's ignore this.

It is even the recommended recipe for simple download.

Update use to tolerate errors in develop mode, to prevent buildout from
deleting develop working copies if anything happens, then this behavior was
extended to non-develop mode, apparently by mistake.

In non-develop mode, we want error to propagate, if something went wrong with
updating this repository, buildout should delete and retry.

The purpose is to replace colletive.recipe.shelloutput

/reviewed-on !11

These options were used by a subclass that does not exist anymore.

This reverts the following commits partially:
  d157d8a1
  e92d31cf

With buildout 2, indentation is either fully removed (if anything
starts on the same line as '<option-name> =') or preverved (the same
number of leading spaces if stripped). Any use that required these
commits was a misuse. They could even break valid Python.

This fixes a problem visible when installing openssl 1.1.1d from
component/openssl. At the end of install step, some symlinks are created
inside openssl install directory for all certificates from
ca-certificates part, then the "make read only" step occurs, which
recursively chmod on files and folder in openssl installed parts, but
since this was following symlinks, it was also trying to make readonly
in ca-certificates parts, which in my case was belonging to another
unix user, because it was a shared part folder from outer level slapos.

/reviewed-on nexedi/slapos.recipe.build!10

It's a nightmare to convert from DOS to GPT (required when moving to a
disk bigger than 2TB), because GRUB can only be installed reliably
from the system to boot.

ext4 was chosen for the 'discard' option, but formatting takes too
much time and space. I was surprised to see my VM being paused
(no left space on host) because of ext4lazyinit.