Targets for (unspecified, soonish):

Integration of smart_holder into refactored code as optional add-on
- #include <pybind11/pybind11.h> will only see is_smart_holder_base_tag but not any other smart_holder code.
- #include <pybind11/smart_holder.h> needed to actually bring in the smart_holder code.
- refactoring of smart_holder code itself, e.g. trim smart_holder_poc.h into a smaller memory/smart_holder.h, move rest to split out tests/memory/smart_holder_poc.h.

Targets for 2.8:

Decoupling of unit tests
- Currently test A cannot run until all tests A-Z are compiled.
  - Blocking in parallel build environments
- Link time with optimization is currently ~63 seconds (gcc 10.2, -Os -flto).
  - Also very long startup times (>15s seconds) in distributed build environments because of very large .so size in debug mode (~100M).
- Tricks needed for fast development-test iterations.
- Fix: minimize use of TEST_SUBMODULE, use only if the submodule feature itself is being tested.
Systematic refactoring
- Unit tests must not be modified (with only very rare, narrow exceptions).
- No change to internals (no exceptions).
- Refactoring into subdirectories.
- Existing header files will mostly be shims, provided for backward compatibility.
  - While we're at it: leave nothing considered private in namespace detail.
Optional precompilation
- WRT PR #2445: jumping between 2 x the number of files is a permanent tax on development activity and an impediment to continuous refactoring.
- Keep code together, use #ifndef PYBIND11_USE_PRECOMPILE_MODE.
- Develop (small-ish) tool to split out .cpp files, integrate with CMake.

Better MyPy integration
Better Scikit-build integration (waiting on Scikit-build developments)
Future of type_caster_base
Multiple-inheritance overhaul (to fix highly unobvious (to pybind11 users) UB)
- Porting Boost.Python inheritance code.
- Will probably be another round of refactoring, much smaller scale, but significant.

Code written here is licensed under pybind11's license (BSD-3) unless otherwise stated.