Bug report

Bug description:

Dear Python team,

I am the lead of thenanobind project, which is a C++<->Python binding tool. I always try to follow recent Python versions to react to any API/ABI changes and catch issues before they end up in public releases. With the just-released Python 3.14a7, a whole bunch of testcases in the nanobind test suite start to fail (example:https://github.com/wjakob/nanobind/actions/runs/14372986244/job/40299387939). These are all tests that create and pass objects and expect reference counts to behave in a certain way. This works consistently for PyPy and all tested CPython version from 3.8 all the way until 3.14a6, and so it is therefore surprising to see such a change in behavior in an alpha version bump.

I'm wondering what could cause this? Are there known issues in 3.14a7? Did reference counting / garbage collection change in some way that could cause such behavior to arise?

Thank you,
Wenzel

CPython versions tested on:

3.14

Operating systems tested on:

macOS, Linux

Linked PRs

• gh-132346: Docs: Clarify that reference counts aren't stable between versions #132352
• [3.14] gh-132346: Docs: Clarify that reference counts aren't stable between versions (GH-132352) #136613
• [3.13] gh-132346: Docs: Clarify that reference counts aren't stable between versions (GH-132352) #136614