[bkrabach]

_configure_context() mutates the wrapped ssl.SSLContext — on Linux it calls
ctx.set_default_verify_paths(), which writes the context's X509_STORE.
wrap_socket() holds self._ctx_lock across the __enter__() for that reason
(#192 / #193, released in 0.10.4). wrap_bio() still does a bare
with _configure_context(self._ctx):.

That gap matters because wrap_bio() is the path async stacks take — anyio →
httpcore → httpx never reach wrap_socket() — so the unlocked path is the one
async clients actually use.

What we saw

An application that mounts a session by resolving ~40 agents' model roles
concurrently (one asyncio.gather, one HTTPS call each) dies partway through
with no Python traceback and no result:

double free or corruption (fasttop)
Fatal Python error: Aborted

Exit 134, sometimes SIGSEGV/139 with the same stack. With
PYTHONFAULTHANDLER=1, every abort captured had the same shape — 20–37 threads
inside _configure_context, all of them via wrap_bio, none via
wrap_socket:

Thread 0x0000e7965cf9f180 (most recent call first):
  File ".../truststore/_openssl.py", line 38 in _configure_context
  File "/usr/lib/python3.12/contextlib.py", line 137 in __enter__
  File ".../truststore/_api.py", line 144 in wrap_bio
  File ".../anyio/_backends/_asyncio.py", line 1100 in run
  File "/usr/lib/python3.12/threading.py", line 1073 in _bootstrap_inner

truststore 0.10.4, Python 3.12.3, OpenSSL 3.0.13, Linux aarch64, in a container.

Verification

A/B/A in one disposable container, this branch built as a wheel and installed
over the stock 0.10.4 — same probe, same concurrency, only truststore changing:

condition	concurrent attempts	sequential attempts	aborts
stock 0.10.4	10	6	2 (plus 1 on the very first, unpaired sequential run)
this branch	20	6	0
reverted to stock 0.10.4	10	—	3

Every abort carried the signature above: 25–31 threads in _configure_context,
100% via wrap_bio, 0 via wrap_socket. An earlier independent run of the same
A/B/A got 7 aborts in 15 stock concurrent attempts (and 1 SIGSEGV in 10 stock
sequential), 0 in 20 patched, 2 in 5 after reverting.

The change

Deliberately the same shape wrap_socket() uses, including holding the lock only
across __enter__() and not for the duration of the wrap — so concurrent wraps
and handshakes are not serialized.

Tests

• test_wrap_bio_configures_context_under_lock — deterministic. Patches
  _configure_context to record whether _ctx_lock is held while the context is
  being configured, and asserts the lock is released again afterwards. It fails
  on main (records False) and passes with this change.
• test_threading_wrap_bio — a threaded hammer over wrap_bio(), in the style
  of the existing test_threading. Worth being honest that this one is only
  probabilistic: without the lock the failure is a native abort that takes the
  whole test process down rather than a failed assert, so the deterministic test
  above is the real regression guard.

Full suite passes locally on Linux (nox -s test, Python 3.12, with mkcert and
internet access): 61 passed, 2 skipped.

Notes

• Cache automatic OpenSSL verify path setup #213 (caching the verify-path setup) would reduce how often this mutation
  happens, but the lock looks worth having either way.
• Race condition on context setting #209 is a different race in the same area — the window after the lock is
  released, where check_hostname/verify_mode are still relaxed. This PR does
  not address that one.
• Happy to add a CHANGELOG entry, drop the stress test, or reshape any of this —
  whatever fits best.

CI note

The three pypy3.10 jobs fail on this branch exactly as they do on main (its last three runs, Sep 3–5): cryptography cannot build its wheel there because PyO3 0.29 requires PyPy ≥ 3.11 — the failure is in pip install -r dev-requirements.txt, before any test runs. Every CPython job (3.10–3.14 on all three OSes), Lint, and the pip integration job pass, including the new regression test.

[sethmlarson]

Thanks for this contribution and the explanation. Could you add an entry to the CHANGELOG?