GeezerVison

Summary

Separate constructor vectorcall from instance-call capability, and distinguish retained raw support from automatic vectorcall eligibility.

• Keep builtin constructor implementations on their own non-inherited slot, with a separate type-instance resolver.
• Inherit instance capability and eligibility under their respective conditions. A generic Python __call__ revokes automatic eligibility without deleting raw support.
• Make the existing PyVectorcall_Call entry point use retained raw support directly, without an artificial tp_call fallback/precondition.
• Recheck eligibility when a retained callable wrapper invokes cached vectorcall support. This handles a partial subclass changing __call__ between two regex replacement calls.
• Update builtin registrations, constructor specialization and partial dispatch consistently. No new test namespace or Lib changes.

This is a Python 3.15 upgrade prerequisite, not the complete _testcapi fixture integration.

Validation

Tested commit 0624705569dff7c3954f9129ba8b9d04469872d7 on base e5554e23abfbaee85508c32fc22e5074f88aebdd, Linux x86-64, Rust 1.99.1:

• CLI build: passed.
• Strict VM Clippy, default and threading: passed.
• VM no-default-features library check: passed.
• Strict derive-impl library Clippy: passed.
• Strict C API library Clippy from its configured crate directory: passed.
• Normal repository commit hooks: passed.

A separately built CPython 3.15.0 final (d220409a13b77de7a6c30abe16ea90db01528edc) was compared with the source-bound RustPython binary in fresh processes:

• 37 constructor/builtin plus 14 partial records: all 51 portable results agree.
• 26 capability/flag records: 25 agree. RustPython's wrapper_descriptor retains its genuine native vectorcall capability, whereas CPython does not expose that flag for its implementation. Immutable-type and method-descriptor flags agree. This is not reported as universal flag parity.
• The bounded two-substitution callback case returns AB on both interpreters after __call__ changes during the first call.
• 19 selected existing canonical test_functools/test_call tests: 19 ordinary passes, zero failures, errors, skips, expected failures or unexpected successes.

No baseline runtime comparison, benchmark, full test suite, external C ABI runtime, PyO3 bridge runtime, or non-Linux execution is claimed. C API validation here is source review plus compilation/linting.

Integration boundaries

Raw storage and automatic eligibility are now distinct. Related call-slot inheritance work in #6301 must preserve that distinction rather than mechanically transplanting the old architecture. For a plain-metaclass adaptation that intentionally disables automatic vectorcall, the current mechanism is slots.flags.remove(PyTypeFlags::HAVE_VECTORCALL), not deleting retained raw storage. This is an unvalidated downstream integration note, not a claim about that PR's broader compatibility.

This patch does not implement arbitrary concurrent native slot replacement, cached-None/ArgCallable redesign, C API kwargs error-precedence changes, or the custom-MRO missing-call/fresh-dispatch edge. It does not change lifecycle ownership or the #8944 retirement contract.

Implementation, review and automated validation were AI-assisted. The model version identifier is unavailable. Automated validation is not human verification.