[mpage]

Acquire coherent strong-reference snapshots under the partial object lock for calls, repr, reduce, and nested-partial construction. Publish replacement state atomically, release displaced references after publication, and retire replaced instance dictionaries through QSBR. Keep free-threaded vectorcall dispatch immutable after construction.

Add concurrent state/coherence coverage, placeholder-aware flattening tests, and regressions for reentrant keyword merging and finalizers.

Authored with AI and reviewed by me.

• Issue: Free-threaded CPython 3.14: concurrent functools.partial.__setstate__ crashes with repr(), calls, or another __setstate__ #157841

[espressolee]

Ran the reproducer from the issue against this PR (cae9916) and its parent (738d055). macOS arm64, --disable-gil builds, 4 s per run.

Release builds, PYTHONMALLOC=debug, 10 runs per mode:

mode   738d055        cae9916
ro     10/10 ok       10/10 ok
rw     0/10 (segv)    10/10 ok
ww     0/10 (segv)    10/10 ok
cw     0/10 (segv)    10/10 ok

TSan builds (--with-thread-sanitizer, in-tree FT suppressions file which has no active entries, halt_on_error=0), 3 runs per mode:

mode   738d055               cae9916
ro     no reports            no reports
rw     reports in 3/3 runs   no reports
ww     reports in 3/3 runs   no reports
cw     reports in 3/3 runs   no reports

The parent had 6-21 data race warnings per rw/ww/cw run, with top frames in partial_setstate and partial_setvectorcall; those runs ended in SIGABRT, SIGKILL or a 300 s timeout. All 12 PR TSan runs exited 0 with no warnings. No crashes or TSan reports with the PR in these runs.