GeezerVison

Summary

Extract the ContextVar cache-identity fix from #8954 for the Python 3.14 main branch.

A ContextVar can retain a cached value after its Context is destroyed. If a new empty Context reuses that address at the same stack depth, the old address-based cache check can return the destroyed Context's value even though the new mapping is empty.

Give every Context, including copies, a non-reused cache identity. If the identity counter saturates, zero disables cache hits. Existing cache locking and replacement/drop ordering stay intact. Add a Rust interpreter regression for fresh and copied contexts, variable defaults, explicit defaults, and missing-value errors.

This changes one implementation file and adds one integration test. Canonical Python tests, dependencies and build profiles are unchanged.

Validation

Independently validated on main-based head 15cb422e0 with its own Python 3.14 library:

• The exact new integration test fails on unpatched main and passes with the fix.
• A batch-allocation probe confirms actual address reuse: unpatched main returns stale values at reused addresses; patched dev and release runs return zero stale values after reuse for both fresh Contexts and copies.
• The initial shorter CLI probe observed no address reuse after the layout change and was retained as inconclusive. Its assertions were not weakened; the separate batch probe exercises the missing case.
• Normal hooks and strict workspace/C-API Clippy passed. Workspace tests: 1,353 passed, 18 existing ignored. Separate C-API tests: 116 passed, 4 existing ignored.
• Dev and production-feature release builds passed. Full unchanged test_context passes in both: 56 tests, one existing skip. Both existing contextvars/threading-contextvars snippets pass, and the release context suite also passes through cargo run.

Local execution was Linux only. Asyncio cohorts requiring the sandbox-denied AF_UNIX socketpair were not run; hosted tests remain enabled and cross-platform CI is pending. Counter exhaustion was source-reviewed without mutating the global counter. The available CPython comparison was 3.15.0rc3, not a CPython 3.14 reference run.

AI assistance

Codex assisted with implementation, extraction, source review, regression tests, and automated validation. The original implementation authorship and Assisted-by trailer are preserved; the regression commit also includes its disclosure.