fix(cmake): carry option-driven defines on pybind11_headers INTERFACE

[henryiii]

This is a general fix anyway, can stand on its own.

🤖 AI text below 🤖

Description

PYBIND11_DISABLE_HANDLE_TYPE_NAME_DEFAULT_IMPLEMENTATION and PYBIND11_SIMPLE_GIL_MANAGEMENT were applied with a directory-scoped add_compile_definitions, so they only took effect in the master-project (test) build. Move them onto the pybind11_headers INTERFACE, the same pattern as PYBIND11_INTERNALS_VERSION. The options now also work in add_subdirectory mode and ride the exported/installed targets.

This is step 0 of the optional pre-compilation work: must-match configuration macros have to propagate through pybind11::headers so a future precompiled static library and consumer modules always agree.

Suggested changelog entry:

• The CMake options PYBIND11_DISABLE_HANDLE_TYPE_NAME_DEFAULT_IMPLEMENTATION and PYBIND11_SIMPLE_GIL_MANAGEMENT now apply in add_subdirectory mode and to installed/exported targets

[rwgk]

Reviewed with codex gpt-5.6-sol ultra:

This change looks right to me: these configuration-dependent definitions belong on pybind11_headers as usage requirements, so they propagate through the target graph and survive export/install.

I'm approving, but I would strongly suggest adding regression coverage before merging. We already have focused CMake integration tests under tests/test_cmake_build, including small add_subdirectory and installed-target consumers, so this should fit naturally into the existing structure.

The current option-enabled CI configurations do not distinguish the old behavior from the new behavior: directory-scoped definitions already reached pybind11’s top-level test targets, while the separately configured consumer tests neither forward these options nor assert the resulting macros.

I think the existing subdirectory_target and installed_target tests could be extended by:

• forwarding the two enabled option values into the nested configurations;
• defining test-only “expect this macro” flags on the consumer targets; and
• adding conditional #ifndef checks to their shared main.cpp.

Existing CI jobs already enable each option and run test_cmake_build, so this should not require new workflow jobs. The resulting tests should fail on master and pass with this PR.

[henryiii]

I'd push back on the test. Adding cmake tests are expensive (one push of this stack kills the ability to run anything on any pybind repo for several hours!), we don't do cmake tests that often, and it will get tested by the items later in the stack anyway.

[henryiii]

Oh, I just realized you have a proposal that extends an existing test, that's a bit more attractive. I can check.

[henryiii]

🤖 AI text below 🤖

Extended the existing subdirectory_target and installed_target tests as suggested: the outer build forwards the two option values into the nested configurations, the consumer targets get EXPECT_<option> defines, and the shared main.cpp errors if the expected macro did not propagate. Verified locally that both tests fail with master's CMakeLists.txt and pass with this PR.