Sphinx 7.2 breaks generated documentation headers

[oddbookworm]

Describe the bug

Updating to sphinx 7.2.0 (or higher, including 7.2.2) breaks our documentation headers at pygame-ce. As an example, here's the current doc page for pygame-ce's draw module, and here's what the doc looks like when I generate locally with 7.2.0 (7.1.2 produces what's on github)

/* Auto generated file: with makeref.py .  Docs go in docs/reST/ref/ . */
#define DOC_DRAW_RECT "rect(surface, color, rect) -> Rect\nrect(surface, color, rect, width=0, border_radius=0, border_top_left_radius=-1, border_top_right_radius=-1, border_bottom_left_radius=-1, border_bottom_right_radius=-1) -> Rect\ndraw a rectangle"

The reST source for that doc is here and here is our makeref.py

How to Reproduce

Create a pygame-ce dev environment (our github wiki has instructions)
Install sphinx 7.2-7.2.2 (all of these versions give broken doc headers)
python setup.py docs --fullgeneration
python setup.py install
You should see an error about undefined references to DOC_****
Install sphinx 7.1.2 (or earlier) and try again. Should work

Environment Information

Platform:              linux; (Linux-6.1.44-1-MANJARO-x86_64-with-glibc2.38)
Python version:        3.11.3 (main, Jun  5 2023, 09:32:32) [GCC 13.1.1 20230429])
Python implementation: CPython
Sphinx version:        7.2.2
Docutils version:      0.20.1
Jinja2 version:        3.1.2
Pygments version:      2.15.1

This is from my local environment, but we also see the issue in our CI runs for MacOS, Windows, Debian/Ubuntu, and ManyLinux

Sphinx extensions

No response

Additional context

No response

[AA-Turner]

Hi Andrew (@oddbookworm),

Is there a smaller reproducer at all than setting up an entire pygame environment please? Ideally a stand-alone Sphinx project, but in effect the fewer moving parts the better!

Thanks,
Adam

[oddbookworm]

Uh, I don't know enough about how it works to do that myself. Lemme ask around to see if someone else knows enough about our workflow to sort that out

[dr0id]

this is on windows:

it looks like somewhere this is joined without using os.path.join or similar (left 7.1.2, right 7.2.2)

But this is probably not the problem. I have two verbose logs, I can't see any crucial difference in it. I would guess that somewhere the processing of :mod: or module:: is broken.

[ankith26]

I did a bit of digging, and noticed that some change in the doctree attribute of the html-page-context event breaks our usage. I'm not sure how much this info helps, but I'll dig further and try to isolate the issue into a minimum reproducer later

[ankith26]

Though I'm not familiar with the sphinx codebase, to me it looks like a regression from #11533
EDIT: ignore this comment, I misread something

[picnixz]

@oddbookworm If you are not able to find a minimal reproducer, can you bisect the culprit commit so that we may have an idea of what has been introduced / removed that may have lead to this issue? Thank you.

[oddbookworm]

@oddbookworm If you are not able to find a minimal reproducer, can you bisect the culprit commit so that we may have an idea of what has been introduced / removed that may have lead to this issue? Thank you.

I was able to do that.
05a14ff -> works fine
97d2c5d -> broken

[AA-Turner]

This is the last blocker for 7.2.3.

A

[AA-Turner]

The issue is caused by these two lines--swapping the order fixes things. I'm continuting to investigate for a proper fix.

sphinx/sphinx/domains/python.py

Lines 1311 to 1313 in 2f6ea14

	# The node order is: index node first, then target node.
	ret.append(inode)
	ret.append(target)

[AA-Turner]

See #11647, which should fix this issue.