EqualifyEverything / equalify-iris

The CI reviewer now flags docs prose that breaks CONTRIBUTING.md's plain-language rules, as a NON-BLOCKING note, on the file set that document binds: `README.md`, anything under `docs/`, `config.example.yaml`'s comments, and the prompts in `agents/`. CONTRIBUTING.md has required concise plain language since `6bb4fbc` (#454) and nothing enforced it; this is the enforcement half, split out because it edits a workflow. The same prompt already forbids flagging style, so that bullet now names this as its one exception and re-states what stays out of scope inside those files: heading style, line length, Markdown layout, and wording the reviewer would have chosen differently. A prompt holding a rule and its own prohibition produces neither reliably. Four floors keep it from becoming a style review: quote the sentence and name the rule it breaks, three instances at most, detail is explicitly not the target, and five of CONTRIBUTING.md's six rules apply. "Shorter over completer" is left out on purpose — whether a document earns its length is not a judgement a diff supports, so it stays the maintainer's call, and that limit is stated in all three files rather than left as a silent narrowing. `test/review-prompt-plain-language.test.ts` (6 tests) pins the reconciliation, not the wording. The file set has one home — CONTRIBUTING.md's Documentation section — and three enforced copies (the prompt, `docs/ci.md` step 4, `BOUND_FILES`), each pinned by set equality so a widening cannot land in one file alone. It reads the prompt GitHub hands the model rather than the file's bytes, because the clause lives in a block scalar where a byte-level grep matches indentation and wraps that are not part of the string. Twelve review rounds, eleven of them with notes: twenty-three notes, all fixed. Eleven were a message that named the wrong cause, so 81 mutations were run and the last twenty-six were graded on WHICH MESSAGE PRINTED rather than on red/green — an assertion that fires with a remedy belonging to a different input costs a reader a run. 76 killed; the 5 survivors are named in the PR body and are the sentence-grading ones this change deliberately leaves to a human. Three rules came out of those rounds and are written into the test: - Two failures over one region are ordered by which remedy subsumes the other. - A failure is composed from what is MISSING and what is EXTRA, not selected by an `else` — an `else` is a catch-all wearing a diagnosis. - A clause attaches only to the branches whose reason it states, and a branch may ask for a deletion only when the state its own advice produces would still fail. Gates: `tsc --noEmit` clean, `npm test` 1698 / 1698, `./test/e2e.sh` ALL ENDPOINTS PASSED, `actionlint` and `shellcheck` green in CI. No runtime code changed, so nothing about the delivered HTML moves. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>