socketmodule.c: Reference and buffer leaks via audit hook failures

[aisk]

Bug report

Bug description:

Summary

Two leak bugs in socketmodule.c triggered when PySys_Audit raises:

1. getaddrinfo (line 6983): idna and/or pstr refs leaked when audit hook raises. ~657 objects leaked per 1000 calls.
2. sock_sendto (line 4810): pbuf Py_buffer not released when audit hook raises.

Reproducer (getaddrinfo leak)

import socket, sys

sys.addaudithook(lambda *a: (_ for _ in ()).throw(RuntimeError("audit")))
before = sys.gettotalrefcount()
for i in range(1000):
    try:
        socket.getaddrinfo("localhost", 80)
    except RuntimeError:
        pass
after = sys.gettotalrefcount()
print(f"Leaked {after - before} objects in 1000 calls")

CPython versions tested on:

CPython main branch

Operating systems tested on:

No response

Linked PRs

• gh-146245: Fix ref/buffer leaks in socketmodule.c on audit hook failure #146247
• gh-146245: Fix reference and buffer leaks via audit hook in socket module #146248
• [3.14] gh-146245: Fix reference and buffer leaks via audit hook in socket module (GH-146248) #146274
• [3.13] gh-146245: Fix reference and buffer leaks via audit hook in socket module (GH-146248) #146275

[aisk]

Hi @kimimgo, I already had a working PR when I created the issue. I created the issue just to get the issue number as a comment in the test.

How did you create your PR within 10 minutes just after the issue was created? And the news entry filename in your PR is Misc/NEWS.d/next/Library/2026-03-21-00-00-00.gh-issue-146245.SockAud.rst, the time part is just zero. In current LLMs, the KV-cache is very important for performance, so most of the LLM agents will just use zero time in system prompts, so I suspect it is generated by LLM.

If so, please read https://devguide.python.org/getting-started/generative-ai/ before contributing.

[kimimgo]

@aisk You're right — I should have checked for an existing PR first. I'll close my PR (#146247) since yours was already in progress.

To be transparent: yes, I used an AI coding assistant (Claude Code) to help analyze and fix the issue. The speed and the 00-00-00 timestamp in the NEWS entry are artifacts of that workflow. I appreciate you pointing out the devguide AI policy — I've read it and will follow it more carefully going forward.

That said, I do review and validate all changes before submitting — the responsibility for the code is mine. In this case though, your PR takes priority. Apologies for the duplicate, and thanks for the thorough bug report!