autodoc: IndexError formatting signature for callable data object with annotated class variable

[AndiEcker]

Describe the bug

call-object is rendered without its doc-string and sphinx-build outputs the following warning on the console (printed at sphinx.ext.autodoc._dynamic/_loader.py, line 145):

WARNING: error while formatting signature for <module>.<call_object>: list assignment index out of range [autodoc]

this bug happens on build when documenting a module containing a callable data object under the following conditions:

• the call-object's class implements __call__
• the class contains at least one annotated class variable
• the instance is documented using a trailing #: variable doc-string

How to Reproduce

Minimal reproducer

from typing import Any


class SigBug:
    class_var: Any

    def __call__(self):
        pass


sig_bug = SigBug()  #: trigger

Document the module using autosummary / automodule.

Expected behavior

sig_bug should be documented without warnings and should show its docstring ("trigger") in the rendered doc/html/pdf.

Actual behavior

doc-string is not visible in the rendered doc/html/pdf and console shows

WARNING: error while formatting signature for module.sig_bug:
list assignment index out of range [autodoc]

Environment Information

- Sphinx 9.1.0
- Python 3.12.12

Sphinx extensions

autodoc

Additional context

Debugging

The exception originates from

sphinx/ext/autodoc/_dynamic/_signatures.py

at line 128

signatures[0] = (args, retann if isinstance(retann, str) else '')

At this point the local variables are

props.obj_type == "data"
args == "()"
retann is None
signatures == []
result == ("()", None)

signatures is empty because data objects intentionally skip signature extraction earlier in _format_signatures(), but autodoc-process-signature still returns a replacement signature.

The code assumes that signatures already contains an element although this is not true for obj_type == "data".

A possible fix would be to append a new signature when signatures is empty instead of assigning to signatures[0].

[tomatotomata]

I traced this to _format_signatures() leaving signatures empty for data objects, then the signature event handler assigning signatures[0]. I was thinking of adding a regression case for a callable data object with an annotated class variable, then guarding the empty-signature path with the smallest change that preserves the existing behavior. The issue looks unclaimed and I do not see an open PR. Would a focused fix along those lines fit Sphinx's expectations?

[jdillard]

#14578 PR has been made and approved, feel free to review if you want.