feat(language): a parameter value that is null or NaN is refused when the data is attached

[FBumann]

Prompt: Then do a PR about it.
(About Fabian's comment on #750: refuse null and NaN in a parameter table at attach, keep inf as a value.)

Note

The following content was generated by AI.

A parameter value that is null or NaN is refused when the data is attached. A missing row is the only gap a parameter has. inf and -inf stay values. Refs #750. It does not close #750.

Method, gate output, alternatives

What changes

• docs/reference/language/declarations.md, under parameters: the rule, next to the dtype table that the consumer also enforces at attach.
• docs/reference/language/absence.md: one sentence that links to the rule, after "A missing parameter row is not absence".
• ParameterDeclaration docstring (src/mathspec/program.py): the consumer that attaches data refuses a null or NaN, as it refuses a column of the wrong dtype.

Why this is language

Take one spec with one frame. If one engine reads a NaN coefficient as 0 and another lets the NaN drop the row, they build two models (what counts as language). The language already reads a null in a where (ParameterDefined: non-null and finite) and already refuses a null in a relation (relations.md). The consumer enforces the check. specsolve needs the matching refusal.

What this does not do

• It does not decide absence: on parameters, the main proposal of a parameter cannot say that a missing row is absent, so a decision supplied as data brings back the rows its mask deleted #750. It settles only the open point "whether a missing row may also arrive as a null": no, only as no row.
• No code raises the refusal. mathspec attaches no data, so no test can fail without it.
• It does not touch default: (a parameter cannot declare a default, so a bound that data prep fills no longer tells a where: which rows the modeller gave #773). A default fills a missing row. A null row is refused before a default applies.
• A binder that drops or fills PyPSA's NaN defaults before attach is outside this repo. A dropped row reads 0 where a read is unguarded, so the binder has to decide per attribute.

Gates

• pixi run test: 2599 passed.
• pixi run docs-build: no issues.
• pixi run lint: all hooks pass except pyrefly, which fails locally on untouched origin/main as well (the yaml stubs are installed but not found). CI is the check for it.
• compile-tex not run: no file under examples/ changed.

🤖 Generated with Claude Code

[read-the-docs-community]

Documentation build overview

📚 mathspec | 🛠️ Build #34830583 | 📁 Comparing 5738a52 against latest (ae4a44c)

  🔍 Preview build  

4 files changed

± changelog/index.html
± reference/program/index.html
± reference/language/absence/index.html
± reference/language/declarations/index.html