memoryview: traverse the view for the collector, and name types as tp_name does - #8608

Two defects, both found while working through what was left after #8553. Comparisons are against CPython 3.14.6.

A cycle running through a view was never collected

Neither memoryview nor the _buffer_wrapper a __buffer__ produces reported what it held while being traversed, so the object a view looks at was invisible to the collector and the cycle was classified as reachable:

class Src:
    def __buffer__(self, flags):
        return memoryview(b'abcdef')

o = Src()
o.keep = memoryview(o)      # o -> memoryview -> wrapper -> o
del o
gc.collect()                # CPython: freed. before: never freed

The same held for a plain view over an exporter that refers back to it. Both types now report what they hold, as memory_traverse and mbuf_traverse do.

This is what test_gc in test_memoryview was covering; the three expectedFailure decorators are removed, and with them the overrides that only existed to carry the decorator.

An unhashable-type message dropped the module from the name

The message took the short name where CPython uses tp_name:

slot_name() already holds what tp_name holds — name() strips everything before the last dot — so this is that accessor at the two sites that build the message. A class defined in Python is still reported by its own name.

Only this message is changed. name() is read in 321 places, and which of them CPython builds from tp_name rather than __name__ is a separate audit.

Tests

extra_tests/snippets/stdlib_gc.py gains five cycle shapes through a view, and builtin_hash.py five type-name cases. Both files pass under CPython 3.14.6 as well.

16 suites run clean locally — memoryview, buffer, gc, weakref, array, bytes, dict, set, collections, io, re, types, descr, exceptions, pickle, struct.

Left out

• pickle.PickleBuffer is still missing. It is reachable only through _pickle, and test_pickle.py reads a successful import _pickle as "the C accelerator is present", then binds dump, dumps, load, loads, Pickler and Unpickler from it in a class body. A _pickle carrying only PickleBuffer therefore breaks that file at import, so the accelerator has to come first.
• Argument errors carry no function name — readinto() argument must be read-write bytes-like object is reported as buffer is not a read-write bytes-like object. ArgumentError::into_exception never receives the name, which is why the arity messages read expected at least 1 arguments, got 0 too. Threading it through is a change to the argument-binding path rather than to any one converter.

🤖 Generated with Claude Code

https://claude.ai/code/session_01P9HewXGX8qcGSccUxGdSPV

Summary by CodeRabbit

• Bug Fixes

  • Improved garbage collection for memoryview objects and custom buffer exporters, helping cyclic references be collected reliably.
  • Corrected unhashable-type error messages so they display the type name accurately.

• Tests

  • Added coverage for garbage collection across sliced, cast, nested, and custom-exported memory views.
  • Added tests verifying unhashable-type message formatting.