Refactor pytest_collection_modify hook

[ev-br]

Refactor the doctest collection stage a bit.

The key simplifications are

• for dev installs, only run pytest on the build-install folder. Its contents matches the installed version, so it should be identical to $ pytest --pyargs scipy for an installed scipy (I think?). This obviates the need for the PY_IGNORE_IMPORTMISMATCH=1 env var --- almost. The remaining need seems to be due to some weirdness in scipy.signal, which indeed has a matching .py and .so modules:

E   _pytest.pathlib.ImportPathMismatchError: ('scipy.signal._spectral', '/home/br/repos/scipy/scipy/build-install/lib/python3.10/site-packages/scipy/signal/_spectral.cpython-310-x86_64-linux-gnu.so', PosixPath('/home/br/repos/scipy/scipy/build-install/lib/python3.10/site-packages/scipy/signal/_spectral.py'))

This probably needs to be looked at on the scipy side.

• The logfile generation can be done with pytest --collect-only ... and some light postprocessing (below the fold)

The full stanza to run this, including pytest ignores, is currently (I know)

$ python dev.py shell
$ cd ~/repos/scipy/scipy
$ PY_IGNORE_IMPORTMISMATCH=1 pytest build-install/lib/python3.10/site-packages/scipy/ --doctest-modules --ignore=build-install/lib/python3.10/site-packages/scipy/interpolate/_interpnd_info.py --ignore=build-install/lib/python3.10/site-packages/scipy/_lib --collect-only

Log posprocessing scripts are below the fold:

$ cat modify_log.py

# convert the result of 
# $ pytest ... --doctest-modules --collect-only
# to simplify comparing to the doctest_.log from doctest-based runs

PACKAGE = "scipy."

if __name__ == "__main__":
    import sys
    if len(sys.argv) != 2:
        print("Usage: modify_log.py LOGFILE")

    fname = sys.argv[1]
    with open(fname, 'r') as inf:
        in_lines = inf.readlines()

    out_lines = []
    for line in in_lines:
        line_ = line.strip()
         
        if line_.startswith("<DoctestItem"):
            out_lines.append(line_[13:-1])

    print("\n".join(sorted(out_lines)))

and

$ cat modify_doctest_log.py 
# strip module names from the doctest.log
# to simplify comparing to the doctest_.log from doctest-based runs

PACKAGE = "scipy."

if __name__ == "__main__":
    import sys
    if len(sys.argv) != 2:
        print("Usage: modify_log.py LOGFILE")

    fname = sys.argv[1]
    with open(fname, 'r') as inf:
        in_lines = inf.readlines()

    out_lines = []
    for j, line in enumerate(in_lines):
        line_ = line.strip()

        if line_.startswith("="):
            continue

        if j < len(in_lines) -1 and in_lines[j+1].startswith("="):
            continue

        out_lines.append(line_)

    print("\n".join(sorted(out_lines)))

With these, the logs agree for the following combination: https://github.com/ev-br/scipy/tree/doctest_plugin + this PR, and scipy/scipy#16391, modulo #106.

[ev-br]

@Sheila-nk @melissawm any chance you could review this PR? Would be much appreciated.

[ev-br]

scipy/scipy#20112 should take care of IGNORE_IMPORTSMISMATCH

[ev-br]

Merging now to keep the ball rolling as I work on test errors after collection is successful. A post-merge review would still be most welcome.