test: every surface construct is held to the program node it lowers to

[FBumann]

Prompt: "Check the latest 4 open issues. CLosable? Otherwise implement them as PRs"

Note

The following content was generated by AI.

Thirteen rows, expression text in and one program node out: sum(by=), at(by=), shift and sum_back are now held to the node they become and every field it arrives with. Closes #212.

Nothing in src/ changes.

What the table adds, and why a whole-node compare

tests/test_program_nodes.py (#201) asks whether each node kind is reached by some file, and Footprint.shapes (#207) reports the same axis. Any mapping of the four constructs onto five node types passes both. Before this, GroupSum, Translate and Window appeared under tests/ only as hand-built entries in the FAN_IN table — constructed by tests, asserted as lowering's output nowhere.

The nodes are frozen dataclasses, so one == per row asserts the kind and over, into, wrap, fill, partition and width together — the fields a partial assertion skips.

The shape claims, each a row of its own:

• sum_back(within=3) is one Window, not three Translates summed — the number of terms must not be readable from data.
• by=[lk] and by=lk land on the identical node.
• by=[lk, lk2] is one GroupSum with paired tuples, not nested groupings.
• offset=lead carries the parameter name, not a resolved node.
• by= on shift/sum_back reaches the node as partition.

The fixture is fixtures.SMALL_MODEL plus a second groupable lookup and an int parameter, local to the module. tests/test_dimensions.py::BASE was the other candidate; sharing it would couple the dim rules' fixture to a table that will grow with the operator set, which is a coupling neither side asked for.

Verification

pixi run ci green on the branch (lint, 772 tests, docs-build, compile-tex).

A green suite proves nothing about a table nobody can break, so four mis-lowerings were introduced in src/math_spec/lowering.py one at a time and the table run against each:

mutation	result
GroupSum swaps into and coordinate	1 failed, 2 passed
Window drops its partition	1 failed, 12 passed
Translate loses the sign of a literal offset	1 failed, 7 passed
at() lowers to a GroupSum	1 failed, 5 passed

Each mutation reverted; the tree here is unmutated.

Not done

The other eight claims #209 lists, and the vocabulary agreement in its comment — separate PRs on the same base, since #212 is only the shape half.