Fixes MOM6-community#51.

convergent_transport asked the caller to name the vertical layer and interface coordinates, then validated the pair by string substitution:

if layer.replace("l", "i") != interface:
    raise ValueError("Inconsistent layer and interface grid variables!")

str.replace rewrites every "l", so the check only worked for stems containing none: lam_l/lam_i and level_l/level_i were rejected outright, while unrelated pairs whose substitution happened to coincide (ml_l/mi_i) were silently accepted.

This PR started as a better check. It ends by deleting the question: both parameters are gone.

Why the arguments were never needed

• The layer coordinate comes along with the transports. It is a dimension of utr/vtr themselves. On the CM4p25 example file umo has dims (time, sigma2_l, yh, xq) and carries sigma2_l with its values, so it is on the output whether or not anyone names it; dsout[layer] = grid._ds[layer] was re-attaching a coordinate that was already there.
• The interface coordinate comes from the grid. It genuinely cannot ride along — it is one point longer than the layer coordinate, so it is a dimension of nothing on the section — but the grid knows it, and knows it without being told which coordinate the transports are resolved over: whichever axis registers their layer dimension at its "center" position names the interface at "outer"/"inner"/"left"/"right".

So the names were never carrying information the call did not already have. _interface_coords reads the interface off the grid, keyed on the data's own dimensions rather than on an axis called "Z" — so a vertical axis under any name is found, and a grid declaring only its horizontal axes (most of them, since sections are only traced horizontally) simply contributes nothing.

The <stem>l/<stem>i convention leaves the source entirely. Names now play no part: the pairing is the grid's, whatever it spells.

API change

layer= and interface= are removed, not deprecated — pre-1.0, and the old layer="z_l" default was already a landmine (it did grid._ds["z_l"], so it raised KeyError on any grid without that variable). All 32 in-tree call sites passed layer= explicitly; every one is updated.

Downstream callers that read the two names off grid.axes["Z"].coords only to hand them straight back — xwmb/budget.py does exactly this — can now delete those two dict entries. That is hdrake/xwmb#44, open as a draft: the two repos are in lockstep, since xwmb depends on sectionate@master, so it merges immediately after this reaches master and not before.

Verification

• The change is a numerical no-op on real data. On the CM4p25 OSNAP-west section, convergent_transport(grid, i_c, j_c) on this head is xr.testing.assert_identical to the previous call passing layer="sigma2_l", interface="sigma2_i" — same transports, same sigma2_l, same sigma2_i.
• Full suite: 76 passed, 0 skipped, 0 failed (xgcm 0.10.1, with the ECCO LLC90 and MOM6-fold data present so nothing skips). The 19 tests covering the old validation are replaced by 11 covering the new behavior: the layer coordinate arriving from the transports; the interface coordinate from the grid; four name pairs following no convention (including MyCenters/MyEdges); a vertical axis not named "Z"; an interface position naming a bare dimension with no coordinate values; transports with no vertical dimension; and TypeError for each removed keyword.
• All five example notebooks re-executed on this head, in one env, after the merge. No numerical output changed anywhere — comparing every textual output against the previously committed ones, the only differences are the version banner (0.4.0rc2.dev1 → 0.4.0rc2.dev3, a hatch-vcs dev string from an editable install of an untagged branch) and warning messages citing this checkout's paths. Notebook 5's overturning streamfunction reproduces bit-for-bit: psi range (Sv): -58.09992975038242 to 53.35870875408872.

Examples

examples/load_example_model_grid.py now declares the CM4p25 grid's vertical axis ('Z': {'center': 'sigma2_l', 'outer': 'sigma2_i'}), so notebooks 2 and 3 still get sigma2_i on their output — from the grid rather than from a string. Declaring it is only possible because of #8: before that fix build_neighbor_maps padded over every axis the grid carried, so a grid with a vertical axis could not be traced at all.

Merge conflict with #8, resolved

#8 is merged, and both branches had committed re-executed notebooks, so all five conflicted. Resolved by taking this branch's copies wholesale (--ours) rather than hand-merging notebook JSON, then re-executing all five on the merge result — so the committed outputs are the ones this branch's code actually produces, not either side's stale copy. Everything else merged cleanly; #8 touches gridutils.py and two test files this branch does not.

_{Drafted with AI assistance (Claude Code). I have read the diff and run the tests, the notebooks, and the no-op check myself.}