[lmoresi]

development cannot read back a split-mesh checkpoint. #640's fix is in the tree but disarmed, and #640 is labelled fixed-in-development, so #805 would have closed it.

The regression

commit	tests/test_0864_split_checkpoint_roundtrip.py
42f93470 — the merge before #756	5 passed
f414da63 — after #756	1 failed

Both built from clean and run locally. The test is level_1/tier_a and reachable from scripts/test.sh via tests/test_08*py, so CI covers it.

#756 (e4d79d71, "Share compact DG1 geometry across XDMF fields") moved the reload payload from the file root into /restart/petsc:

before   topologies/uw_mesh/dms/vRT/section
         topologies/uw_mesh/dms/vRT/vecs/vRT

after    restart/petsc/topologies/uw_mesh/dms/vRT/section
         restart/petsc/topologies/uw_mesh/dms/vRT/vecs/vRT

Nothing was dropped — I checked the written file, and the per-field section and its global vector are both still there, one group deeper. Only the reader's existence check stayed at the root, so for a file written with petsc_reload=True it answered "no PETSc-native payload". The #640 guard then refused the good path along with the bad:

RuntimeError: read_timestep: the saved field 'vRT' holds 15 coordinates
carrying more than one DOF ... The file carries no PETSc-native payload:
re-write it with Mesh.write_timestep(..., petsc_reload=True).

— advising the flag the caller had already passed.

The fix

Both changes are reader-side, in what the guard is given:

• _petsc_payload_name searches both layouts. It is only an existence check, and a file written by either version has to stay readable.
• _guard_coincident_dofs no longer assumes a v1.1 file has no payload. It looked the payload up for legacy files only, while read_timestep calls _read_native_payload on a v1.1 file immediately afterwards — so the guard refused the files the native path was about to read correctly. The saved-vs-live row comparison stays legacy-only: v1.1 shares one compact geometry across fields, so there is no per-field saved cloud and saved_rows is 0 by construction.

The test that was missing

test_0864 caught the symptom; nothing pinned the cause, so a relocation presented as a confusing refusal instead of a located failure. test_the_payload_is_found_wherever_the_writer_put_it asserts the section and the vector are on disk and that the existence check finds them.

Tier B because it is proven, not assumed: with the single-location lookup restored it fails, and so does the round-trip test; with the fix both pass.

Verification

test_0864_split_checkpoint_roundtrip	6 passed
+ test_0005, test_0007, test_0008, test_0012, test_0020 (checkpoint and snapshot)	50 passed
check_test_coverage.py	all 352 files reachable

How this got in

#756's own CI was green, and its run was 51m46s against roughly 1h30m for its neighbours — worth a look at whether its test job was collecting the full set. The general lesson is the #828 one again: the question is not "do the new tests pass" but does CI exercise the configuration this change affects — here, a file-layout change with the reader's expectations written down in a different file.

Issue state: #640 stays open and fixed-in-development is accurate again once this merges. It should not be closed by the release sweep until then.

Underworld development team with AI support from Claude Code

🤖 Generated with Claude Code

https://claude.ai/code/session_01E87Q7KrpapxeQiLD1RiNXv

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.