fix(language): a long chain of named expressions loads, a typo beside a formal is refused, and a fault in an entry hides no other

[FBumann]

Prompt: Please adress this review

Note

The following content was generated by AI.

Every should-fix and nit of the stack review of #626 through #638 that is a code change, fixed on top of #638, each with the test that fails without it. Stacked on #638. The three #628 findings are fixed in #628 itself, and the stack above it is rebased.

What this changes

Finding	Was	Is
#628 blocking: a formal along= beside by=/within= refused as "no key column over"	fixed silently at #638	fixed in #628's second commit, and a call-site test here asserts the node the call builds
#628: sum(x, by=nope, over=a, into=b) loaded with formal columns	the formals sent the call back before the relation was read	the relation is read first, in #628's second commit
#628 nit: over=/by= refusals lost "or a formal of this macro"	—	the words are back, in #628's second commit
#632: a fault in a macro or an entry hid every other fault	raise after the entries	every declaration is read; a use of a refused entry says it does not load, and the refusal is listed with it
#632: a cycle through a macro read a -> a	expansion knew the macro, the resolver did not	expansion asks the namespace, so the chain is a -> m -> a
#632: a long chain was a RecursionError or the parser's "nests deeper"	each entry recursed into the next	entries resolve from a worklist in dependency order; the resolved tree is held to MAX_RESOLVED_DEPTH (300), and the refusal names that depth
#632 rule: a cycle through a case's when	refused, untested	pinned
#638: unused templates and entries now refused for an amount, an edge or a bare sum	unclaimed	pinned, and claimed in #638's body
#638: a dim fault under a bare sum() is a SchemaError	tests loosened to LanguageError	the class is parametrized per rule; resolution's faults are SchemaError, the dim rules' DimensionError
#638: sum(sum(p, into=g)) added "already a scalar"	the refused inner call was built	a refused call builds nothing
#638: 123456789 < p rewrote to p > 1.23457e+08	:g	the number as the file wrote it
#638: an undeclared piecewise over: added per-link lines	—	one line
#638: "the expansion cannot fail"	a str/bool breakpoint parameter and an lp x-link with no variable were refused under cost_curve_increasing / cost_curve_domain_lo	refused on the block's own link at load
#638 nit: offset=-tag lost the dtype hint	the dtype rule was the dim checker's	a named amount is held to dtype: int where it is read, before its sign
#638 nit: six guards with no test	—	a mutation sweep found nine; three were redundant with the check beside them and are gone, the rest are pinned (below)
#638 nits	+p untested; fan_in crashed on Named; _operands repeated children; the otherwise test at _arms always true; UNREACHABLE named a line that runs	each fixed

Not changed: Mask.names_read (#631 nit) is moot at #638, where both sides of a comparison are Expression; _check_expression's ceiling=None path (#632 nit) is live at #638, where _named passes it.

Why

The review found them, and confirmed each by loading the same model on both branches. I reproduced every one at #638's head before changing anything.

Method, gate output, alternatives

The rebase

#628 gained one commit with the three fixes and their five tests, which fail on its first commit. #629, #631, #632, #638 and this branch were rebased onto it in that order, one lease each. Two conflicts: #632's commit and this one both append to tests/test_expansion.py, both kept; #638's fourth commit rewrote the two functions #628's fix touches, resolved with #638's version and the fix carried into it, so the fix holds at every rung. The suite ran at every rung.

The chain of named expressions

Measured before the change, load then lower then typeset, on a chain e{i} = e{i-1} + 1 read by one constraint:

Declared	#632 head	#638 head	here
deepest last, n=100	ok	ok	ok
deepest first, n=100	RecursionError	ok	ok
n=200	ok / RecursionError	ok / RecursionError	ok
n=250	RecursionError	RecursionError	refused at e150, "nests 301 deep … past the 300 levels"

So the base branch loads 200 only when the deepest entry is declared last. The worklist makes loading order-free; what remains is every later pass recursing over the resolved tree. MAX_RESOLVED_DEPTH = 3 * MAX_DEPTH is where the passes were measured to survive: with the cap off, a chain of 200 survives every pass and 250 does not. docs/reference/language/expressions.md states the limit. The 80-entry chain test_a_call_expands_to_core_ast pins still loads.

Error class

The rule already on main: a dim rule the resolver needs to build a node (a where side's dims, a pullback's landing) is collected with the resolver's faults as a SchemaError. #638 added the bare sum to that set. tests/test_dimensions.py now pins the class per case, and tests/fixtures.py's helpers raise SchemaError as to_spec does.

Guards

A mutation sweep replaced each if of _call, _built, _bare_sum, _edge_fits, _amount, _edge, _dim_ref, _relation_ref, _known_roles, _partition, build, cycle, curve_frame, _piecewise_references and _domain_decides_nothing with if False: in turn and ran the suite: 78 guards, 69 red on the first run. Of the nine green:

• three were redundant, and are gone: a formal amount is caught under its sign and again where a bare name resolves to nothing; a role that named no column is caught where the roles are counted;
• over=1, by='lk' and over=1 as a column had no test: test_a_kwarg_that_names_nothing_is_refused;
• the three frame refusals of curve_frame, and the values-carries-over, points-undeclared and points-carries-over rules of _piecewise_references, had none: seven cases on test_a_malformed_block_is_refused.

The sweep re-run on the touched functions is all red.

Gates

• ruff check, ruff format --check: green. pyrefly check: 0 errors.
• pytest -q -n auto: 1598 passed, 5 skipped (the typst compiles), the walk line census included.
• python -m tools.schema and python -m tests.typesetting.golden: no diff.
• prettier --check on the two changed pages: green. The Read the Docs build passed.
• docs-build did not run here: the session proxy refuses docs.python.org. compile-tex did not run: no tectonic here.

Not done

The bodies of #626, #628, #632 and #638 carried the review's claim findings; each is edited to say what is true now.

🤖 Generated with Claude Code

https://claude.ai/code/session_015h57WkBDnpxrknuJ5zZy9F

[read-the-docs-community]

Documentation build overview

📚 math-spec | 🛠️ Build #34716597 | 📁 Comparing ce399f8 against latest (fdb22f9)

  🔍 Preview build  

23 files changed · ± 23 modified

± Modified

• contributing/index.html
• reference/reading/index.html
• reference/language/errors/index.html
• reference/language/expressions/index.html
• reference/language/named/index.html
• reference/math_spec/degree/index.html
• reference/math_spec/dimensions/index.html
• reference/math_spec/errors/index.html
• reference/math_spec/expansion/index.html
• reference/math_spec/lowering/index.html
• and 13 more...