Restore interrupted with-enter coverage with pending exception support - #9497

Summary

Restore the missing test_with.InterruptDuringEnter regression from pinned CPython d220409a13b77de7a6c30abe16ea90db01528edc through the real update_lib workflow. The canonical class/import are restored unchanged; existing test markers remain unchanged.

The fixture needs genuine deferred delivery rather than an immediate KeyboardInterrupt from __enter__. This adds:

• A VM-owned pending exception-class reference, delivered at bytecode safepoints after existing signals and scheduled GC. Validation/normalization calls the class normally, including its metaclass and constructor, and preserves normalization failure notes.
• Separate generic replacement/notification primitives so the native fixture preserves CPython's exchange → notify → release-old-reference order. The API is restricted to the active VM; this is not cross-thread PyThreadState_SetAsyncExc support.
• The native SelfInterruptingContextManager fixture and explicit root clearing on supported terminal thread/interpreter paths. Reusable enter/run scopes do not clear pending state.

Lifecycle and API review points

• Terminal owners close even an empty slot; the closed state is the last VM field. Existing safe atomic-reference replacement is retained, without new locks or unsafe blocks.
• Normal owner cleanup and matched GILState release clear in an active VM. Unmatched GILState TLS teardown, foreign-thread finalization, already-shutting-down slots, and unavailable slot TLS suppress Python callbacks while releasing this new root instead of attempting an unsafe attachment. This defensive policy is not a guarantee for arbitrary teardown misuse or all existing VM roots.
• Adding Drop to Interpreter restricts moving its public global_state field out directly. In-repository uses were checked; external consumers were not. ThreadedVirtualMachine.vm is not a public external field.
• Pending/closed flags are thread-owned Cell<bool> state on the non-Sync VM. A future cross-thread API needs a separate synchronization/wakeup design. There is still an extra plain flag load per opcode; no performance-neutrality claim is made.

Validation

Local validation used base c2bb450316db3ce3dc4ddc6bcde9f3059ee4e90c. Main advanced to 0072412e9523b37f7c7254c9ceb3468d5041c454 (#9495) during publication; that newer merge composition is left to PR CI, not claimed as locally retested.

The local build/runtime checks below cover original implementation commit 770484ecfd47e6b3af502cc72ce7819fa16e594e: checks preceded that metadata-only commit, and normal hooks left all 3,779 build-bound source files unchanged.

Review follow-up 33de790f50417c21c6a67d3a5a3a0661a53d86ba only changes the class attribute to #[pyclass(name, module = "_testcapi")]. The shared macro name resolver returns the same string from the Rust identifier as from the previous literal; registration and type metadata are unchanged by source inspection. Normal hooks passed. Local compilation/runtime were not repeated for that one-line follow-up; its own PR CI is tracked separately.

On the original implementation source tree:

• Normal commit hooks passed; no hooks bypassed.
• Offline/locked cargo check, strict cargo clippy -- -D warnings, and CLI build passed with threading,stdlib,stdio,importlib,host_env,doc,ucd_3_2_0 and actual VM opt-level 3.
• A separate offline/locked check passed with the same feature list minus threading; the actual VM compiler invocation had no threading cfg.
• The unchanged canonical InterruptDuringEnter.test_exit_called_after_interrupt ran once against the candidate and once against a fresh pinned CPython build containing the real _testinternalcapi fixture. Both ran exactly one test with zero failures, errors, skips, expected failures, or unexpected successes. Inputs were hash-bound, output matched, caches stayed empty, and owned process groups were cleaned up.

This was a bounded Linux single-method comparison, not full test_with, workspace/C-API suites, cross-platform runtime testing, a benchmark, or validation of every generic normalization/finalizer/TLS corner case. No baseline crash probes, OS-signal injection, or cross-thread stress tests were run.

There is file-level overlap with #9143 in _testinternalcapi.rs and VM registration-area files, but this does not duplicate its _testcapi registration work. #9407's C-API tp_new work remains separate. Neither owner's branch was modified.

AI assistance

Implementation, source analysis, independent agent review, guarded validation, and this description were assisted by OpenAI Codex (model version unavailable).