[Culaii]

Closes #387. Supersedes #620 (@override only), which this covers.

Problem

A decorated function is never mutated, except one with a lone @staticmethod / @classmethod. @property, @contextmanager, @override, @abstractmethod, registration decorators (@app.get(...), a plugin registry) all take the function out of mutation testing silently: the score stays at 100% while those bodies are never measured. On the codebase I work on, 318 of 2,814 functions on the mutation surface were skipped this way, 73 of them with real branches.

Change

The approach proposed in @Otto-AA in #387 (7 Mar. 2026): the decorators stay on the trampoline only, and the copies (the original and every mutant) keep only the decorators that decide how they bind -- staticmethod / classmethod, which the dict dispatch needs to call them alike. Against the three reasons in the source comment:

1. A decorator with side effects (@app.post("/foo"), a registry) runs once, on the trampoline, exactly as without mutmut.
2. Decorator nodes are still skipped, so their arguments are never mutated.
3. @property sees the trampoline, a plain function, never a copy.

is_classmethod is detected among all decorators (it was only checked for a lone one), so @deco @classmethod still binds to the subclass.

Still skipped:

• decorators after which another function of the same name follows in the scope -- @overload, @x.setter / .getter / .deleter, @f.register -- since their copies would collide in the mangled names;
• caching decorators (cache, lru_cache, cached_property). Measured on e2e_projects/my_lib: with the cache on the trampoline, all 9 mutants of cached_fibonacci survive although test_my_lib kills most of them -- the parent process fills the cache during the stats run, and every forked mutant answers from it. A per-copy cache (decorating the copies instead of the trampoline) would fix that, but changes caching semantics; I left it out to keep this reviewable and am happy to follow up if you want it.

Tests

• test_decorated_bodies_are_mutated_but_never_their_decorators replaces test_do_not_mutate_top_level_decorators: decorator lines intact, bodies mutated.
• Namesake and caching decorators are not mutated; the copies keep only the binding decorators.
• Runtime (exec) tests: a @property stays a property and its mutants run; @contextmanager; a registering decorator runs once and registers the trampoline, which reaches the active mutant; @deco @classmethod binds to the subclass; an @override method.
• e2e snapshot: the only new entries are Point.skip_multi_decorator's 8 mutants, all killed (the fixture's skip_ name no longer describes it -- happy to rename it here if you prefer).

Suite: 473 passed / 4 skipped on main, 490 / 4 with the change; pre-commit run --all-files passes.

I also ran a backport of the patch on 3.8.0 against ~10 modules of a production codebase: about 72 previously hidden mutants were measured and killed, nothing broke, and the 4 survivors were all genuine (3 equivalent mutants from a redundant condition, 1 missing test).