[harbinresearcher]

Falsey subclasses of the public AdapterFactory were silently skipped by wrapt.decorator, leaving the wrapped function's original signature in place. Check for an adapter factory before evaluating the adapter's truth value, while retaining the existing behavior for None, False, and empty-string adapters.

Add regressions for factories using __bool__ or __len__, checking invocation, signature, argspec, and normal calls. Add compatibility cases for disabled adapters and an unreleased change note.

Validation:

• Both new factory cases fail on the original code.
• Linux: the complete just test matrix passes (27 runs across Python 3.9–3.15, including 3.14t/3.15t, with all three extension variants and the pinned mypy).
• Windows: full core suite passes on Python 3.9–3.14; Python 3.12 full tests including pinned mypy: 1394 passed, 8 skipped.
• Black checks, strict Sphinx build, source/wheel build, and package metadata checks pass.
• Independent agent review found no blocking issues.

Linux used a local Zig compiler adapter because GCC was unavailable; it filters the unsupported GNU --exclude-libs,ALL linker flag and isolates the compiler's Python environment. No project configuration or dependencies were changed. PyPy was not checked locally. AI assistance was used for implementation and verification.

[GrahamDumpleton]

I'll have a look, but be warned that AdpaterFactory and adapter argument to decorator is now deprecated. Instead look at with_signature and with_doc decorators. Can you provide an example of what you were trying to do so I can determine whether the new way of doing things will suit your requirements.

• https://wrapt.readthedocs.io/en/master/decorators.html#signature-changing-decorators
• https://wrapt.readthedocs.io/en/master/bundled.html#signature-override
• https://wrapt.readthedocs.io/en/master/bundled.html#docstring-override

Also see the "Changing the signature" workshop as part of the set of workshops for wrapt.

• https://grahamdumpleton.me/workshops/wrapt/

[GrahamDumpleton]

Thanks for finding this and for the thorough write up. The bug is real: an
AdapterFactory instance which evaluates as false was being skipped because
the truth value of the adapter was tested before checking whether it was a
factory.

I have applied a fix directly on develop in 77e826d rather than merging
this PR, for two reasons.

First, I restructured the conditional. Rather than testing
isinstance(adapter, AdapterFactory) or adapter and then repeating the
isinstance check inside, the factory is now resolved by type first and the
truth value of the result is tested afterwards. That removes the duplicated
check and reads as the intended logic rather than a patch over it.

Second, that restructuring changes one additional behaviour. A factory which
returns None, or another false value, now means no adapter is to be used for
that function. Previously it fell into the argspec handling and failed with
an obscure TypeError. The change note and tests cover this new case as well
as the ones from this PR.

The tests here were adapted into the commit, and the change note sits under
the 2.5.1 section with credit to you. You are also recorded as a co-author
on the commit. Closing this in favour of that commit.