Mark wrapper descriptors as method descriptors - #9286

Summary

Mark wrapper_descriptor as METHOD_DESCRIPTOR, matching CPython 3.15.0. This is the isolated descriptor-flag hunk from the Python 3.15 upgrade work, not the wrapper lifetime/GC changes tracked separately in #8944.

The flag allows existing method lookup paths to retain the descriptor and prepend the receiver rather than allocate a bound wrapper. The existing direct, bound and vectorcall paths already perform the same receiver validation and slot argument dispatch. The only Lib change removes one obsolete RustPython expected-failure marker from test_monitoring.TestLoadSuperAttr.test_vs_other_type_call, regenerated through update_lib from pinned CPython 3.15.0. All canonical assertions and other local adaptations are preserved.

Validation

• Fresh dedicated-target CLI build with freeze-stdlib passed. All 3,631 tracked source/symlink hashes matched before and after the build.
• Two selected unchanged canonical methods passed on main baseline, candidate and pinned CPython 3.15.0: test_descr.OperatorsTest.test_lists and test_descr.ClassPropertiesAndMethods.test_pow_wrapper_error_messages.
• One unchanged flag assertion extracted from canonical test_call failed on baseline and passed on candidate and CPython. This is an external observation, not execution of the CAPI-dependent enclosing test or canonical-suite coverage.
• Candidate observation process completed within the fixed 15-second, sampled 512 MiB RSS, 6 GiB address-space and core-zero bounds.
• Threaded VM check (170s), non-threaded VM check (46s), strict affected VM all-target Clippy (118s), and normal hooks (4s) passed.

These narrow controls do not establish correctness of every optimized dispatch path or a performance improvement. Full workspace/CAPI runtime gates and the release-mode whole-module command have not been run locally and remain outstanding; actual hosted executed steps will be tracked separately. The full CAPI-dependent method and whole test modules were not executed locally.

Monitoring CI repair

The original head 0af373cf failed hosted Ubuntu tests because the now-correct test_vs_other_type_call unexpectedly succeeded, including the workflow's built-in rerun. On the exact unchanged canonical selector, the sealed main baseline fails because int.__repr__ CALL/C_RETURN events report MISSING instead of receiver 0; the sealed candidate and pinned CPython both pass with zero skips. The updater-generated correction at 87db988c removes exactly that expected-failure decorator, with no native change or assertion weakening. All three local runs used the fixed 15-second / 512 MiB sampled-RSS guard; normal hooks passed. These are one-selector results, not a local full-module or full-suite claim. Fresh hosted CI is pending.

AI assistance

This change and its source review/validation were prepared with OpenAI Codex assistance. No human review is claimed by this disclosure.

Performance qualification for measured head 0af373cf

CodSpeed executed the benchmark and reported a passing gate, but comprehension_list.py increased from 803.995 to 891.509 µs (−9.8164% efficiency). The report exposes RELATIVE detection, allowedRegression=0.1, confidenceLevel=0.99, and no benchmark-specific override. Its NoChange classification therefore does not establish equal performance.

The comparison uses main bcfad3554ef66264d08c2d030d1fa13aaab5c03c and PR head 0af373cff4c5b603a3d723f28d3fcb5ba8f6c5e8, tested as merge 8c101451f39dd8d8e18c4bf393c0ed6f73aded0c. Both named environment fingerprints are ffdf350d53eb784b, but the baseline job used kernel 7.0.0-1017-azure and runner image 20261018.339.1, whereas the head used kernel 6.17.0-1022-azure and image 20261012.334.1. Both used CodSpeed runner 5.4.2. This is not a fully matched-environment comparison and does not establish a cause. Four other benchmarks were skipped with baseline results reused.