…readed use

native LOG() routes records through Python's logging, so it acquires the GIL
from C++. several sites did that while holding a native mutex (the
connection-pool mutexes, the per-connection child-handles mutex, the logger's
own mutex) or the getEnvHandle static-init guard. under concurrent use with
DEBUG logging on, that inverts lock order against a thread that holds the GIL
and is waiting on the same native lock, so the process deadlocks at 0% cpu.

this makes native logging never hold a native lock across a GIL acquisition:
build the log values under the lock, release the lock, then log; construct the
Connection and close pools outside the pool mutexes; release the GIL around the
env-handle init.

verified no deadlock at 1/2/4/8/12/16 threads with logging on (was 100%
deadlock at 4 threads). logging output and the pooling/logging/stress test
suites are unchanged.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

…logging-deadlock

…ault run

removes the standalone tests/_issue671_deadlock_workload.py script (an odd
non-test file in tests/) and moves the workload into a __main__ block in
test_025, which the tests invoke as the child process. one self-contained file,
matching the layout of the other test modules.

also drops the default (non-stress) test to 2 threads to match the repo's
lightest concurrency baseline (test_020's (2, 50)); 2 threads is issue #671's
stated minimum and still deadlocks the pre-fix build 3/3 runs, so the guard
stays reliable without loading PR validation. the heavy 16-thread variant stays
under @pytest.mark.stress. log file now uses pytest's tmp_path (auto-cleaned).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

…est spawn

Two follow-ups from the PR review on the #671 deadlock fix.

connection.cpp: Connection::disconnect() emitted LOG()/LOG_ERROR()
unconditionally, but the same function already refuses to log on the GIL-less
destructor/shutdown path (LOG() acquires the GIL internally, which can hang or
std::terminate during interpreter shutdown). Gate every log site in disconnect()
on hasGil, computed once up front, and make hasGil robust after Py_Finalize()
by short-circuiting on Py_IsInitialized() before PyGILState_Check().

tests/test_025: forward the parent sys.path to the child subprocess verbatim
(os.pathsep.join(sys.path)) instead of dropping empty entries. An empty entry
means the current working directory; filtering it can leave a script-launched
child unable to import the same local mssql_python. Matches test_023.

Regression tests:
- test_child_pythonpath_forwards_cwd_entry: asserts the empty (cwd) sys.path
  entry survives forwarding; fails if the old filtering is reintroduced.
- test_debug_logging_pooled_connection_shutdown_exits_cleanly: DEBUG-logged
  pooled-connection teardown at interpreter shutdown exits cleanly (a teardown
  smoke guard; the pooling atexit handler drains the pool before finalization,
  so the pure GIL-less static-destructor branch is not deterministically
  reachable from Python).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>