docs(limits): a whole-table operator is priced at a barrier rather than refused

[FBumann]

Prompt: "We decided to NOT add an escape island!!" and "the whole table read is not a wall... Its a performance loss. Nothing severe"

Note

The following content was generated by AI.

What this changes

Two sentences on limits.md were wrong, and both are corrected here.

A whole-table operator is priced, not refused. The page said no engine could build such a model one chunk of rows at a time. What it actually costs is one full pass: the engine reads the whole table before it builds any chunk, and the chunks stop being independent. An operator worth that enters, and a request for one names the price. Self-reference keeps its refusal on its own ground — nothing bounds how far it expands — so the operator table now splits the two apart.

There is no escape hatch. #38 was closed not_planned on 2026-08-27 and the docs never followed. The refusal table's second row was no engine could build it one chunk of rows at a time, with a whole-table operator and arbitrary Python as its examples. The operator has left that row, so the row is now the file would stop being the artifact — arbitrary Python, whose content no loader can check and no typesetter can print — and its answer to "can it change?" is no. errors.md replaces the #38 paragraph with the rule: the language has no escape hatch, and a gap closes as a macro, a primitive or a formulation.

AGENTS.md follows — a primitive is admissible when it is relational, and locality prices it rather than barring it.

Why

Both sentences are load-bearing, and a reader acts on them. The first told anyone proposing an operator that the engine could not build it, which is an impossibility claim the engine does not make. The second advertised a hatch that had already been decided against, so a reader planning around it was planning around nothing.

The choice to open the door rather than keep the refusal on a corrected reason was taken in the conversation that prompted this. Locality becomes what a primitive is priced against, not a gate it has to pass.

What was verified, and what was not

pixi.sh is blocked from this session's network, so pixi run ci was not run. In its place:

• prettier --check clean across docs/ and AGENTS.md. It reflowed two tables in limits.md; both reflows are in the diff.
• tests/test_docs.py -k "generated or generator" — 14 passed. None of the three changed files is generated, and the six generated pages are untouched.
• Full tests/test_docs.py shows 10 failures under an ad-hoc uv environment. The same 10 fail on a stashed clean tree, so they are that environment's mkdocs plugin set and not this change. A real pixi run ci is still owed before merge.

Prose measured with the docs-writing skill's script: limits.md avg 18.7 → 17.9, median 16, 20 sentences over 25 words before and after and all of them pre-existing; errors.md avg 17.2 → 16.7, median 14, over-25 5 → 4. No sentence this change adds exceeds 25 words.

Deliberately not done: no src/ change. The language admits no whole-table operator today and this PR does not add one — it corrects what the page says the bar is. Admitting an actual operator is a separate request, judged against the rule as now written.

🤖 Generated with Claude Code

https://claude.ai/code/session_01T18AdG3hoawcwuqXxr4dnM

[read-the-docs-community]

Documentation build overview

📚 math-spec | 🛠️ Build #34566077 | 📁 Comparing 87df947 against latest (78b0ce1)

  🔍 Preview build  

2 files changed

± about/limits/index.html
± reference/language/errors/index.html