test: a program node nothing produces is caught, rather than waiting for a consumer to meet it

[FBumann]

Prompt: "Add a PR onto main that fixes that!"

Note

The following content was generated by AI.

cases: reached a release candidate in #168 with no lowering at all — CasesNode joined the AST, nothing produced a program node for it, and the suite was green on 780 tests. assert_never catches that only where some test happens to lower a file using the construct, and none did.

This is the fence that would have caught it on the first commit:

AssertionError: every_program_node.yaml lowers to none of ['Unlowered']. A node no file reaches is
one whose lowering nobody has run — add a declaration using the construct it stands for.

It is the sibling of test_the_golden_model_carries_every_node_kind_the_walk_renders, one state along: that one holds the renderer to the AST, this holds the lowering to the program.

The two fixtures cannot be one

The obvious move is to point this at tests/typesetting/golden/model.yaml, which already exists to carry every construct. It does not lower:

constraint 'composed': shift() over a variable-free expression leaves vacated positions with no
value, and inventing one is what silently pinned a bound to zero.

That is correct, not a bug — rendering accepts more than lowering does, so a fixture proving every render arm is reached cannot also prove every lowering arm is. tests/fixtures/every_program_node.yaml is the second one, and it is a file that lowers to all thirteen nodes and nothing else.

examples/ cannot serve either: the gallery reaches twelve of thirteen, missing Power. Adding a ** to a teaching model to satisfy a test would be the harness showing through the modelling.

Both directions

The reverse assertion — the fixture lowers to nothing that is not a program node — is what stops the fixture drifting into asserting nothing, the way five fences went vacuous at the language cut by scanning a directory that no longer existed.

Mutation table

Taken by hand — tools/mutate.py cannot express an added class — with a clean tree, git checkout -- to restore and __pycache__ dropped on both sides.

Mutation	Caught
a node joins ExpressionNode with nothing producing it	✅ names it, and says what to do

That is the whole point of the fence, so it is the whole table: the mutation reproduces the exact shape cases: shipped in.

Verified separately, before writing it, that the fence is not vacuous: run over examples/*.yaml it goes red on main today, missing Power.

Verified: pixi run ci — lint, tests, docs-build --strict, 26 TeX documents. The fixture lives under tests/ and is not a gallery model, so nothing renders it.

Not done: the same fence for Predicate. The program's masks are the language's own WhereNodes since #177, so they are already held by the where-parser's tests — there is no second set to drift. Worth revisiting only if the program ever grows a mask type of its own.

[read-the-docs-community]

Documentation build overview

📚 math-spec | 🛠️ Build #34273860 | 📁 Comparing 1b638d6 against latest (58e1d35)

  🔍 Preview build  

1 file changed

± CHANGELOG/index.html