Track sectionate 0.4.0rc2: drop the removed layer/interface kwargs

[hdrake]

Makes topology-overhaul work against the pre-released sectionate >= 0.4.0rc2.

What changed and why

sectionate 0.4.0rc2 removed the layer and interface keyword arguments from convergent_transport:

rc1: (grid, i_c, j_c, f_c, utr, vtr, layer, interface, outname, sect_coord, geometry, positive_in, cell_widths)
rc2: (grid, i_c, j_c, f_c, utr, vtr,                  outname, sect_coord, geometry, positive_in, cell_widths)

Upstream that is hdrake/sectionate#7, "Take the vertical coordinates from the data and the grid; drop layer/interface". The reasoning there: the layer (cell-center) coordinate is a dimension of utr/vtr themselves, so it already rides along into the section — naming it only re-attached a coordinate that was on the output either way. The interface coordinate genuinely cannot ride along (it is one point longer, so it is a dimension of nothing on the section), but the grid knows it without being told: whichever axis registers the transports' layer dimension at its "center" position names the interface at "outer"/"inner"/"left"/"right". So neither parameter is needed, both were removed rather than re-validated, and passing either now raises TypeError.

regionate's library code never passed them — only the real-grid tests and the example notebooks did. This PR bumps the floor and updates those call sites.

File	Change
pyproject.toml	sectionate >= 0.4.0rc1 → >= 0.4.0rc2
regionate/tests/test_real_grids.py	3 call sites lose layer=/interface=
examples/1..5*.ipynb	the same 8 notebook call sites, re-executed
.github/workflows/ci.yml	drop the dead git-ref sectionate install + its stale TODO

No assertion, tolerance, or substantive check was changed. The tests still verify exactly what they did before (the discrete divergence theorem, the Arctic-fold single-region trace, the ECCO hard-topology closures).

The one behavioral difference, and why it is inert

Two of the three test call sites passed layer=None, which was already a no-op. The third passed layer="z_l", interface="z_i" on a grid that declares only X and Y axes. Under rc2 I confirmed empirically on that exact grid:

• z_l still arrives on the section output, with the right values ([3000]), because it is a dimension of T_adx/T_ady. The test's .sum("z_l") is unaffected and the summed value is unchanged.
• z_i is no longer attached, because no axis of that grid registers z_l at "center" for _interface_coords to read the interface name off.

Nothing in regionate or its examples reads z_i — it is only ever created, in the two data-setup blocks. So the loss has no consumer and no number moved. If we later want z_i to keep riding along, the upstream-intended fix is to declare the vertical axis on the xgcm.Grid ("Z": {"center": "z_l", "outer": "z_i"}) rather than to name it at the call site; I did not do that here because it is a separate change with no current consumer. Flagging it as a judgement call worth a second opinion.

The pre-release floor resolves without --pre

The CI comment asserts this, so I checked it rather than assuming. A clean venv (pip 26.2.1), pip install --dry-run -e ., no --pre:

Would install ... sectionate-0.4.0rc2 ... xgcm-0.10.1 ...

A pre-release floor permits pre-releases of that package, so neither --pre nor a git ref is needed in CI.

Test counts actually measured

Clean venv, Python 3.12.7, sectionate==0.4.0rc2, xgcm==0.10.1, xarray 2026.7.0, regionmask 0.13.0:

Suite	Before (at 5e4a17b, on rc2)	After
pytest -q	48 passed, 10 skipped	48 passed, 10 skipped
REGIONATE_REALDATA_TESTS=1 pytest -q	50 passed, 8 failed	58 passed

Every one of the 8 failures was TypeError: convergent_transport() got an unexpected keyword argument 'layer'. Both suites were re-confirmed on a second clean venv at Python 3.13.9 with the same pins: 48 passed / 10 skipped and 58 passed.

For reference, the same commit on sectionate==0.4.0rc1 gave 48/10 and 58 passed — i.e. rc1 was already green, and this PR restores that on rc2.

Notebooks: re-executed

Yes — re-executed, in a clean venv pinned to the authoritative versions (sectionate==0.4.0rc2, xgcm==0.10.1), on Python 3.13.9 with matplotlib==3.10.3 and cartopy==0.25.0 chosen to match the toolchain the committed outputs were produced with. All five ran end to end with no errors.

I diffed every human-readable output (stdout streams and text/plain results) cell by cell, before vs after. Every substantive line is byte-identical — including Check whether the discrete Divergence Theorem holds: [[ True]] (nb 1) and divergence-theorem residual = 0.00e+00 (nb 5). Dropping the kwargs moved no number anywhere in the examples.

Exactly two textual lines differ, both machine-specific paths inside warnings, neither related to this change:

1. 3_Arctic_heat_CM4p25.ipynb — the xgcm experimental north-fold UserWarning prints the emitting file's absolute path. It changes from your …/envs/regionate-pr22/… to my clean venv's path under /private/tmp/…/scratchpad/nb313/…. You will probably want to regenerate this one cell locally so the docs don't carry a scratch path; I left the genuine output rather than hand-editing it.
2. 4_bounded_by_named_sections.ipynb — a tight_layout warning quoting the ipykernel_<pid> temp path, which changes on any re-execution on any machine.

The plot images also churn: PNG payloads shift slightly (different font/DPI metrics on this machine) even with matplotlib pinned to the same 3.10.3. That is unavoidable re-execution noise, not a change in what is plotted.

What I chose not to do

• No new test. Per CLAUDE.md norm 4 I considered adding one, but this change touches only test call sites, notebook call sites, and dependency metadata — there is no new library behavior to cover, and the real-data suite going 50/8-failed → 58-passed on rc2 is the test. Inventing a test that asserts convergent_transport no longer accepts layer= would be testing sectionate's API from the wrong repo; upstream already covers it (TypeError for both removed keywords).
• Did not declare a Z axis on the fold test's grid or on examples/load_example_model_grid.py's construct_grid, which would restore the z_i coordinate on section output. See above — no consumer, and it is a separate decision.
• Did not touch docs/requirements.txt; it had already dropped its git-ref line.

Uncertainty a reviewer should resolve

• Whether the scratch-venv path now embedded in 3_Arctic_heat_CM4p25.ipynb's warning output is acceptable, or should be regenerated locally.
• Whether losing the z_i coordinate on section output is fine (I believe it is — nothing reads it) or whether the grids should declare their vertical axis.
• CI itself is unverified from here: the workflow only triggers on main, so this PR into topology-overhaul will not run it. My evidence for the CI edit is the local dry-run resolution above, not a green run.

🤖 Generated with Claude Code

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[hdrake]

Looks great!