fix: bind noexcept and ref-qualified methods from unregistered base classes

[Skylion007]

Description

Fix issue #2234 . Fix overloading to accept noexcept function types too and strip them. Mostly authored by ClaudeCode with a lot of human intervention.

fix: support noexcept and ref-qualified methods from unregistered base classes (issue #2234)

Problem

In C++17, noexcept became part of the function type (P0012R1). This means
int (Base::*)() const noexcept and int (Base::*)() const are distinct
types. When a derived class inherits methods from an unregistered base,
method_adaptor must recast the member function pointer from Base::* to
Derived::*. Without explicit template specializations for noexcept (and
ref-qualified) variants, the generic F&& fallback was selected, leaving the
pointer typed as Base::* — causing pybind11 to look up Base as self,
which fails at runtime for unregistered base classes.

Similarly, cpp_function lacked constructors for noexcept, & noexcept,
const noexcept, const & noexcept, &&, const &&, && noexcept, and
const && noexcept member function pointer types, so those overloads silently
fell through to the generic lambda constructor.

Solution

include/pybind11/pybind11.h

• Add cpp_function constructors for all missing qualifier combinations:
  • &&, const && (unconditional): use std::move(*c).*f in the lambda
  • noexcept, const noexcept, & noexcept, const & noexcept,
    && noexcept, const && noexcept (under #ifdef __cpp_noexcept_function_type)
• Add detail::rebind_member_ptr<Derived, T> type trait mapping
  Return (Base::*)(Args...) qualifiers to Return (Derived::*)(Args...) qualifiers
  across all 12 qualifier combinations, generated via a local
  PYBIND11_REBIND_MEMBER_PTR macro.
• Add detail::adapt_member_ptr<Derived>(pmf) — shared constexpr helper
  called by all method_adaptor overloads: runs the is_accessible_base_of
  static_assert then returns the implicitly-cast pointer.
• Replace the ad-hoc method_adaptor overloads with a PYBIND11_METHOD_ADAPTOR
  macro covering all 11 qualified variants (plus the no-qualifier overload
  written explicitly to avoid MSVC C4003). All overloads are constexpr.
• Add unconditional & / const & method_adaptor overloads (no #ifdef
  guard) to fix C++14 builds where noexcept is stripped from the type but
  the & qualifier is not.

include/pybind11/numpy.h

• Add vectorize overloads for noexcept free function pointers.

Tests

• test_methods_and_attributes: NoexceptDerived / NoexceptUnregisteredBase
  (C++17 noexcept from unregistered base), RValueRefDerived /
  RValueRefUnregisteredBase (&&/const && methods that move-out state to
  confirm std::move(*c).*f semantics), NoexceptOverloaded (overload_cast
  with noexcept), static_asserts verifying method_adaptor preserves all
  qualifier combinations.
• test_buffers: def_buffer with noexcept member function pointers.
• test_numpy_vectorize: vectorize with noexcept free function pointers.

Compatibility

• No behaviour change for existing code — all new cpp_function constructors
  and method_adaptor overloads are additive and selected only for the new
  type signatures.
• C++14 compatible (noexcept constructors/overloads gated on
  __cpp_noexcept_function_type).
• MSVC C4003 avoided by writing the no-qualifier specializations explicitly
  rather than invoking macros with an empty argument.
• clang-tidy bugprone-macro-parentheses suppressed with NOLINTBEGIN /
  NOLINTEND around the macro definitions (the qualifiers argument appears
  in type position where parenthesizing it would be invalid C++).

Suggested changelog entry:

• Fix binding of noexcept and ref-qualified (&, &&) methods
  inherited from unregistered base classes in C++17, where noexcept is
  part of the function type. method_adaptor now has specializations for
  all qualifier combinations (const, &, const &, &&,
  const &&, and their noexcept variants), and cpp_function gains
  matching constructors. &&-qualified methods correctly use
  std::move(*self).*f in the generated lambda.

[Skylion007]

Should I normalize it here for all pybind11? Or keep it as utility that and create a new alias that always strips the noexcept

[Skylion007]

on the flip side we could use the noexcept tag handling to optimize away exception handling for certain edge cases like def_buffer maybe?

[Skylion007]

Also unsure if should encourage rvalue overloads, but they were always supported by manually calling them via lambdas sooo

[rwgk]

Hi @Skylion007 could you please validate my high-level understanding of this PR?

This is the result of me chatting with Cursor (GPT-5.3 Codex High) for a bit:

---

• Status quo on current master is inconsistent/partial.

  • Some noexcept callables bind today (often simpler signatures).
  • Others fail to bind cleanly, or fall through less-specific template paths.
  • In some cases (notably around unregistered-base/member-pointer adaptation), binding may compile but not behave correctly at runtime.

• Throwing through bound noexcept functions remains process-terminating by C++ semantics.

  • If a C++ exception escapes a noexcept function, std::terminate is triggered.
  • This preempts pybind11 exception translation (error_already_set -> Python exception).
  • Therefore, Python-visible calls may abort the process if a bound noexcept function throws.

• What PR 5992 improves.

  • It broadens/normalizes qualifier handling (const, &, &&, and noexcept combinations), making signature matching more uniform and predictable.
  • It reduces accidental fallback paths and fixes specific C++17 noexcept-typed member-pointer adaptation issues (issue noexcept base class methods are not recognized in C++17 mode #2234 context).
  • It does not change the underlying C++ runtime rule for exceptions escaping noexcept.

[rwgk]

Hi @Skylion007 I added five Cursor-generated commits (all using GPT-5.3 Codex High). I'll do a manual review here after I see that the CI is still passing.