fix: bump PYBIND11_INTERNALS_VERSION to 13

[henryiii]

🤖 AI text below 🤖

Description

Closes #6090.

Bumps PYBIND11_INTERNALS_VERSION to 13 before the v3.1.0 release. #5960 added noexcept to the hash functors in internals.h. With libstdc++, a noexcept hasher disables per-node hash caching in unordered containers, which changes the node layout of inactive_override_cache. Modules built from master before and after #5960 both claim internals v12 but disagree on that layout, so v12 is not safe to ship as-is. A comment now marks the noexcept as ABI-relevant to prevent an accidental change later.

Investigation of #6090 (partially revert the noexcept additions) shows the revert itself is not necessary:

• The custom string-based type_hash / type_equal_to only compile under libc++ (_LIBCPP_VERSION); libstdc++ and MSVC use std::hash<std::type_index>. libc++ stores the hash in every node unconditionally, so noexcept has no effect there — a benchmark of both variants under libc++ shows identical node size (32 bytes) and identical insert/find/rehash times.
• override_hash is the one functor where noexcept changes libstdc++ behavior (24- vs 32-byte nodes). It hashes two pointers, so recomputation is nearly free; the smaller uncached node is the better trade. Keep the noexcept.

For reference, a benchmark of the general caching trade-off under libstdc++ (GCC 16, djb2 string hash, 2000 keys, 4M lookups) shows caching only matters for expensive string hashes: hits 148 ms vs 115 ms, misses 257 ms vs 147 ms, 200 rehashes 9.9 ms vs 0.8 ms (noexcept/uncached vs cached). That configuration does not occur in pybind11, which is why no revert is needed.

Suggested changelog entry:

• Bump PYBIND11_INTERNALS_VERSION to 13, because internals: optimize std::unordered_map internals with noexcept #5960 changed the internal override-cache node layout with libstdc++.

[rwgk]

Hi @henryiii, two thoughts:

• I see you closed this already
• I'm thinking it's indeed not needed, because I batched three PRs that I merged in quick succession:
  • Expand float and complex strict mode to allow ints and ints/float (for PEP 484 compatibility). #5879
  • Improve performance of enum_ operators by going back to specific implementation #5887
  • internals: optimize std::unordered_map internals with noexcept #5960

Did you discover that already, and is that why you closed this PR?

[henryiii]

Haha, no, this was supposed to be #6090, the removal of noexcept. But between asking it to make the PR, and it making the PR, Claude noticed that this path could only be hit with libc++, and libc++ don't change behavior like libstdc++ does. The initial experiments we were doing to set up the PR was intentionally targeting libstdc++ with a MWE, which did show that it would affect libstdc++. Since it noticed that, it dropped the changes, but left the version bump and pushed the PR with only that before I could stop it.

(It isn't needed due to the quick succession, but that's why it kept that change and opened the PR anyway even after removing the real change.)

[Skylion007]

Yeah this is the one I wanted to remove, but it's a relatively minor nit.

[henryiii]

It was a very valid nit if this was not scoped to only libc++.

[Skylion007]

Yeah I guess the question is for very long cstrs it could technically be a performance regression, but probably fine