[rkent]

Description

The noexcept specializations added in #3322 make the rosdoc2 documentation build fail: Sphinx's C++ domain hits an internal AssertionError while resolving references in the function_traits specializations. Wrap the noexcept variants in \cond/\endcond so Doxygen skips them. They mirror the documented non-noexcept specializations, and the compiled code is unchanged.

Is this user-facing behavior change?

No, affects documentation only.

Did you use Generative AI?

Yes, Opus 5.5

More details as reported by Opus:

These lines are there to keep Doxygen from documenting the two noexcept specializations. Without them, the rosdoc2 build of rclcpp crashes in Sphinx.

What breaks without it

The unfixed run (runs_rolling/rclcpp/log.txt) gets through Doxygen. Then Sphinx dies while resolving cross-references:

File ".../sphinx/domains/cpp/init.py", line 1115, in _resolve_xref_inner
assert parent_symbol # should be there
AssertionError
The patched run (runs_rolling/fix_rclcpp/log.txt) ends with build succeeded.

Why noexcept is the trigger

In C++17, noexcept became part of a function's type. So R(Args...) and R(Args...) noexcept are different types, and function_traits needs separate partial specializations for each. That's the point of the comment in the code.
Doxygen sees four separate specializations and outputs four separate class pages. The log shows entries like function_traits< ReturnTypeT(Args ...) noexcept >.
Breathe/Exhale turns each page into a Sphinx cpp:struct declaration. Sphinx's C++ domain doesn't handle noexcept inside a function type used as a template argument. In its symbol table, the noexcept specialization ends up colliding with (or failing to attach to) the plain one. When Sphinx later resolves a reference to it, the parent symbol it expects isn't there, and an internal assert fails. Because it's an assert, it doesn't just log a warning like the "Duplicate C++ declaration" messages elsewhere in the log; it aborts the whole build.
To be clear about the evidence: the traceback and the before/after runs are confirmed. The exact way Sphinx mishandles noexcept is my best reading. The log has no warning naming these structs before the assert.

Why \cond / \endcond

/// \cond … /// \endcond is Doxygen's way to skip a block. Everything between the markers is left out of the docs, but the compiler still sees it, so the code behaves exactly the same. The only loss is two doc pages nobody needs: these specializations just inherit from the non-noexcept versions, which are still documented.

It's the same idea as the DOXYGEN_ONLY trick a few lines further down in function_traits.hpp. That one shows Doxygen a simpler declaration; this one hides the problem declarations completely.

[mjcarroll]

Just a documentation change and passes Rpr, does not need full CI.

[github-actions]

ABI Compliance Check

✅ Verdict: compatible

Library	Verdict	Summary
libcomponent_manager.so	✅ compatible	No ABI changes detected.
librclcpp.so	✅ compatible	No ABI changes detected.
librclcpp_action.so	✅ compatible	No ABI changes detected.
librclcpp_lifecycle.so	✅ compatible	No ABI changes detected.

✅ libcomponent_manager.so — full abidiff report

Compared:

• Base: lib-base/libcomponent_manager.so
• Head: lib-pr/libcomponent_manager.so @ 0a1311c

(empty report — no differences printed by abidiff)

✅ librclcpp.so — full abidiff report

Compared:

• Base: lib-base/librclcpp.so
• Head: lib-pr/librclcpp.so @ 0a1311c

(empty report — no differences printed by abidiff)

✅ librclcpp_action.so — full abidiff report

Compared:

• Base: lib-base/librclcpp_action.so
• Head: lib-pr/librclcpp_action.so @ 0a1311c

(empty report — no differences printed by abidiff)

✅ librclcpp_lifecycle.so — full abidiff report

Compared:

• Base: lib-base/librclcpp_lifecycle.so
• Head: lib-pr/librclcpp_lifecycle.so @ 0a1311c

(empty report — no differences printed by abidiff)

_{Updated for commit 0a1311c · check run · suppressions: /home/runner/work/_temp/ros2-abi-suppressions.txt}