wraps(): concurrent or reentrant calls share and clear one statistics dict, violating thread-local contract

Summary

BaseRetrying.wraps() binds the same wrapped_f.statistics dict object into every per-call retry copy (copy._local.statistics = stats) after clearing it. When the decorated function is invoked from two threads concurrently — or re-enters itself within one call stack — both invocations share, clear, and mutate one unlocked dict. Statistics from one call are destroyed mid-run by another, contradicting the documented contract that statistics are local to each call/thread.

Static analysis of tenacity/__init__.py on current master; not executed by the reporter.

Minimal Reproduction (static)

import threading
from tenacity import retry, stop_after_attempt

@retry(stop=stop_after_attempt(3))
def flaky():
    ...  # fails a couple of times

threads = [threading.Thread(target=flaky) for _ in range(2)]
[t.start() for t in threads]
[t.join() for t in threads]

print(flaky.statistics)
# Expected: stats reflecting this invocation's own run
# Actual: whichever thread cleared/mutated `stats` last wins;
#         attempt_number / start_time / idle_for are cross-contaminated

Location

tenacity/__init__.py, BaseRetrying.wraps → inner wrapped_f:

copy = self.copy()
# Reuse the same statistics dict rather than rebinding the attribute
# so that the stats stay visible through additional decorators that
# copy attributes via functools.wraps (which copies the reference to
# this dict into the outer wrapper's __dict__). See issue #519.
stats = wrapped_f.statistics  # type: ignore[attr-defined]
stats.clear()
copy._local.statistics = stats  # noqa: SLF001
self._local.statistics = stats
return copy(f, *args, **kw)

Why This Is a Bug

1. The docstring/attribute contract states statistics values "are local to the thread running call" (see the comment block on RetryCallState / statistics ~25 lines above). Sharing and clearing a single dict across concurrent invocations breaks that guarantee.
2. The #519 fix solved real visibility problem (outer decorators seeing stats through functools.wraps), but the implementation makes the wrapper-level dict a global rendezvous point: every concurrent entry point calls stats.clear() while another invocation is mid-retry-loop, so:
   • attempt_number can reset to 0 between attempts,
   • start_time reflects the wrong invocation,
   • idle_for/delay_since_first_attempt accumulate across unrelated runs.
3. Recursive/re-entrant use (f calling itself through the same decorated wrapper) hits the identical clobbering single-threaded: the outer run's statistics are wiped when the inner call returns.

The pre-#519 behavior (fresh state via self.copy() without rebinding) was race-free but broke decorator-chaining visibility; the current code fixes visibility at the cost of isolation. Both properties should be satisfiable simultaneously.

Suggested Direction

Keep publishing the final RetryCallState into wrapped_f.statistics after the invocation completes (preserving #519 visibility), but let each execution build its own state inside copy._local during the run:

copy = self.copy()
try:
    return copy(f, *args, **kw)
finally:
    wrapped_f.statistics.clear()
    wrapped_f.statistics.update(copy.statistics)

This preserves single-final-value semantics for readers while eliminating mid-run cross-thread mutation. A test spawning two threads against a multi-attempt decorated function and asserting each sees its own attempt_number progression would lock the behavior in.

Environment

• tenacity master (post-#658), file tenacity/__init__.py
• Python 3.x (threading from stdlib)