[antonsoo]

A copied body with a nonzero center of gravity initially reports the correct position, then jumps on its first position update:

body = pymunk.Body(2, 9)
body.center_of_gravity = (2, 3)
body.position = (10, 11)
restored = body.copy()
pymunk.Body.update_position(restored, 0.1)
print(restored.position)  # currently Vec2d(8.0, 8.0), despite zero velocity

Restore center_of_gravity before position, following the maintainer's suggested attribute ordering. Older saved states still carry the old ordering, so restoration also puts position last in a fresh attribute list. Position is assigned once, and the caller's saved state is unchanged.

Regression checks cover dynamic/kinematic bodies, deepcopy and pickle protocols 2 and 5, then perform a native position update. Two added subtests load the historical attribute order; the simple ordering-only variant fails those, and the final correction passes them. Current upstream now includes the independently merged segment-neighbor correction in #319. Full native tests and mypy pass on the rebased correction; the native Munk2D source and reused extension are unchanged.

A native continuation check uses Space.copy and pickle protocols 2, 4 and 5, three seeds, eight/32 bodies, offset/centered bodies, gravity and angular motion. All 48 corrected histories exactly match their uninterrupted source histories and independent center-of-mass equations. Another 36 cases load frozen archives written by unchanged current upstream: 18 displaced scenes and 86,400 wrong body-step observations become zero. All frozen input hashes and all original histories remain exact. Newly written pickles use the requested attribute order; their raw hashes consequently change. These are contact-free physics scenes; this does not address the documented collision-cache limitation.

[viblo]

Thanks for finding the bug and the PR, great to have someone looking into the copy/pickle feature!

For the PR, I think instead of setting the position twice you could rearrange the order of the attributes in the _pickle_attrs_general so that cog is set before position. Then position would not need to be set twice.

[antonsoo]

Thanks. Updated the attribute order and removed the second position assignment. Older pickles retain their saved attribute order, so restoration also puts position last without changing the caller's state; two regressions cover that case. All 48 current-copy scenes and 36 frozen-archive scenes match uninterrupted motion exactly. The branch is rebased onto the #319 merge.

[viblo]

Thanks!
(A minor note, I updated the code after merge to not fix unpickling of pickles made with old versions. To keep the code cleaner for this special edge case.)

[antonsoo]

Understood. #321 removes the old-version state test and retains the current-format coverage. It is rebased onto the #320 merge.