paperlint 2.1.0 → 3.0.0

package/docs/rules.md

@@ -3,29 +3,63 @@
 The README lists what each check catches in one line. This page says exactly which file each one
 reads and when it fails. Errors fail `paperlint lint`; warnings print and do not.
 
-| Rule                           | Level | Reads                                         | Fails when                                                                                                                                                                                                                                                                                                                                                                                                                                                      |
-| ------------------------------ | ----- | --------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `paper/stages`                 | error | `PIPELINE-STATUS.md`                          | a declared stage's PDF is missing, or the `bytes:` it names does not match the file's real size, or a frozen PDF exists that no stage declares                                                                                                                                                                                                                                                                                                                  |
-| `paper/source`                 | error | `PIPELINE-STATUS.md`                          | a declared stage has no frozen `.tex` beside its PDF (a commit hash does not count — squash and gc destroy it)                                                                                                                                                                                                                                                                                                                                                  |
-| `paper/author-list`            | warn  | `PIPELINE-STATUS.md`                          | a stage is declared and **no cell of the scorecard's table contains the run marker** (default `bib-authors`). It checks that you recorded the cross-check; it does not read your `.bib`. What to run is your own `authorListCommand`, empty by default                                                                                                                                                                                                          |
-| `paper/research-question`      | warn  | `paper.tex`, `paper.md`, `PIPELINE-STATUS.md` | a stage is declared and either the scorecard has no `researchQuestion` field, or it has one the paper does not contain. The declared sentence is compared against the source with whitespace collapsed — your words, not a pattern. Advisory, because nothing here can judge whether what you declared _is_ a research question                                                                                                                                 |
-| `paper/typography`             | warn  | `paper.tex`, `paper.md`                       | any of four counts rises above the per-paper allowance you set: `§` or `\S\ref` instead of "Section"; a decimal with no leading zero (`.05` instead of `0.05`, IEEE / ISO 80000-1 style), counted only where the reader sees it — prose, math and table cells, not macro options, column specs, comments, code or tikz; `Fig.` and `Figure` mixed in one document; bibliography entries with no doi, url or arXiv id. Existing debt is tolerated, growth is not |
-| `tex/future-promise`           | warn  | `paper.tex`                                   | a camera-ready build still says "will be released" about something already handed over                                                                                                                                                                                                                                                                                                                                                                          |
-| `tex/acm-frontmatter-override` | error | `paper.tex`                                   | an `acmart` build overrides ACM's front-matter commands and drops template elements from page 1                                                                                                                                                                                                                                                                                                                                                                 |
-| `review/findings-cause`        | error | `reviews/*.md`                                | a review lists at least `minFindings` (default 3) findings and no cell introduces a cause with the marker (default `Cause:`)                                                                                                                                                                                                                                                                                                                                    |
-| `doc/fields`                   | warn  | `reviews/*.md`                                | a front-matter field is missing or holds a value outside the list you configured. Off entirely unless you configure `docFields`                                                                                                                                                                                                                                                                                                                                 |
-| `pdf/fresh`                    | error | `paper.tex` → `_build/paper.facts.json`       | the paper names a venue and the facts cannot be judged: not JSON, a schema other than 2, or they describe a PDF that is gone or differs from the one on disk (its SHA-256)                                                                                                                                                                                                                                                                                      |
-| `pdf/profile`                  | error | `paper.tex` → `paperlint.json`                | `paperlint.json` is not JSON or has an unknown key, only a pre-2.1.0 `venue.json` is there (`npx paperlint init` moves it), its `extends` does not resolve (not found, a cycle, a chain longer than four, a preset that fails the schema, an npm name), it names no `kind` while the preset has kinds, or names a kind the preset does not have                                                                                                                 |
-| `pdf/fonts`                    | error | `paper.tex` → `_build/paper.facts.json`       | a font the pages draw is Type 3 or not embedded, or no font starts with the family the venue preset names for body text (`fonts_text`) or headings (`fonts_title`)                                                                                                                                                                                                                                                                                              |
-| `pdf/geometry`                 | error | `paper.tex` → `_build/paper.facts.json`       | the page width or height is more than `dimTol` (default 0.05 in) off the preset's, or the column count differs                                                                                                                                                                                                                                                                                                                                                  |
-| `pdf/limits`                   | error | `paper.tex` → `_build/paper.facts.json`       | body or reference pages exceed the limit of the paper's kind, or the reference font size is outside the preset's range widened by `body_pt_tol`                                                                                                                                                                                                                                                                                                                 |
-| `pdf/body-size`                | warn  | `paper.tex` → `_build/paper.facts.json`       | the body font size is more than `body_pt_tol` off the preset's. A warning: banal measures the mode of the rendered text, not the declared size (9.30 pt measured at a declared 9)                                                                                                                                                                                                                                                                               |
-| `pdf/measured`                 | warn  | `paper.tex` → `_build/paper.facts.json`       | the paper's `paperlint.json` names no venue preset yet; or it names one and there are no facts (it was not built), or the facts carry no page geometry (banal was not found) — so the checks above did not run                                                                                                                                                                                                                                                  |
+| Rule                           | Level | Reads                                         | Fails when                                                                                                                                                                                                                                                                                                                      |
+| ------------------------------ | ----- | --------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `paper/stages`                 | error | `PIPELINE-STATUS.md`                          | a declared stage's PDF is missing, or the `bytes:` it names does not match the file's real size, or a frozen PDF exists that no stage declares                                                                                                                                                                                  |
+| `paper/source`                 | error | `PIPELINE-STATUS.md`                          | a declared stage has no frozen `.tex` beside its PDF (a commit hash does not count — squash and gc destroy it)                                                                                                                                                                                                                  |
+| `paper/author-list`            | error | `paper.tex` → `_build/references.json`        | an entry's authors are not those of the version it cites — typically the arXiv preprint's list under a published venue. Checked online by `paperlint build` (DBLP), reported on the entry's own line                                                                                                                            |
+| `paper/research-question`      | warn  | `paper.tex`, `paper.md`, `PIPELINE-STATUS.md` | a stage is declared and either the scorecard has no `researchQuestion` field, or it has one the paper does not contain. The declared sentence is compared against the source with whitespace collapsed — your words, not a pattern. Advisory, because nothing here can judge whether what you declared _is_ a research question |
+| `paper/section-word`           | warn  | `paper.tex`, `paper.md`                       | `§` or `\S\ref` instead of the word Section. **Fixable:** `--fix` writes `Section~\ref{…}` / `Section 5`                                                                                                                                                                                                                        |
+| `paper/leading-zero`           | warn  | `paper.tex`, `paper.md`                       | a decimal below 1 without its leading zero (`.05` instead of `0.05`, IEEE / ISO 80000-1 style), counted only where the reader sees it — prose, math and table cells, not macro options, column specs, comments, code or tikz. **Fixable**                                                                                       |
+| `paper/figure-ref-style`       | warn  | `paper.tex`                                   | `Fig.~\ref` and `Figure~\ref` mixed in one document; each minority occurrence is reported. **Fixable:** `--fix` writes the majority form                                                                                                                                                                                        |
+| `bib/reachable-entry`          | warn  | `paper.tex` (its `filecontents` bibliography) | an entry has no doi, url or arXiv id — a reader has nothing to follow. Reported on the entry; no fix, the link has to be looked up                                                                                                                                                                                              |
+| `paper/cite-exists`            | error | `paper.tex` → `_build/references.json`        | an entry's doi or arXiv id resolves to nothing, or to a different work. Checked online by `paperlint build`; a work that was merely not found is not a finding                                                                                                                                                                  |
+| `paper/refs-fresh`             | error | `paper.tex` → `_build/references.json`        | the bibliography changed since the build checked it (its SHA-256 differs) — run `npx paperlint build`                                                                                                                                                                                                                           |
+| `paper/refs-checked`           | warn  | `paper.tex` → `_build/references.json`        | the paper has a bibliography and no build has checked it, or the last build could not (no network) — so the two rules above did not run                                                                                                                                                                                         |
+| `tex/future-promise`           | warn  | `paper.tex`                                   | a camera-ready build still says "will be released" about something already handed over                                                                                                                                                                                                                                          |
+| `tex/acm-frontmatter-override` | error | `paper.tex`                                   | an `acmart` build overrides ACM's front-matter commands and drops template elements from page 1                                                                                                                                                                                                                                 |
+| `review/frontmatter`           | error | `reviews/*.md`                                | the review's frontmatter fails paperlint's JSON Schema — a `findings` record without `id`/`status`, an unknown field, or an **open** finding with no `cause` (`skill-defect` · `missing-skill` · `hook` · `rule`). A review with no `findings` key is not checked for findings                                                  |
+| `sibling/frontmatter`          | warn  | `siblings/*.md` (not `README.md`)             | a sibling card's frontmatter has no `read:` saying how much of the competing paper was read (`full` · `abstract` · `none`)                                                                                                                                                                                                      |
+| `pdf/fresh`                    | error | `paper.tex` → `_build/paper.facts.json`       | the paper names a venue and the facts cannot be judged: not JSON, a schema other than 2, or they describe a PDF that is gone or differs from the one on disk (its SHA-256)                                                                                                                                                      |
+| `pdf/profile`                  | error | `paper.tex` → `paperlint.json`                | `paperlint.json` is not JSON or has an unknown key, its `extends` does not resolve (not found, a cycle, a chain longer than four, a preset that fails the schema, an npm name), it names no `kind` while the preset has kinds, or names a kind the preset does not have                                                         |
+| `pdf/fonts`                    | error | `paper.tex` → `_build/paper.facts.json`       | a font the pages draw is Type 3 or not embedded, or no font starts with the family the venue preset names for body text (`fonts_text`) or headings (`fonts_title`)                                                                                                                                                              |
+| `pdf/geometry`                 | error | `paper.tex` → `_build/paper.facts.json`       | the page width or height is more than `dimTol` (default 0.05 in) off the preset's, or the column count differs                                                                                                                                                                                                                  |
+| `pdf/limits`                   | error | `paper.tex` → `_build/paper.facts.json`       | body or reference pages exceed the limit of the paper's kind, or the reference font size is outside the preset's range widened by `body_pt_tol`                                                                                                                                                                                 |
+| `pdf/body-size`                | warn  | `paper.tex` → `_build/paper.facts.json`       | the body font size is more than `body_pt_tol` off the preset's. A warning: banal measures the mode of the rendered text, not the declared size (9.30 pt measured at a declared 9)                                                                                                                                               |
+| `pdf/measured`                 | warn  | `paper.tex` → `_build/paper.facts.json`       | the paper's `paperlint.json` names no venue preset yet; or it names one and there are no facts (it was not built), or the facts carry no page geometry (banal was not found) — so the checks above did not run                                                                                                                  |
 
 Optional rules — off unless you turn them on in the `rules` setting, because only some venues need
 them — are on their own page: [`optional-rules.md`](optional-rules.md). Today there is one,
 `pdf/last-page-balance`.
 
+## A deliberate exception: a disable directive, with the reason
+
+When a finding is right in general and wrong for one line — a quoted `.05`, a bibliography entry
+for an invited talk that has no link — keep the exception where it is, with ESLint's own directive
+in a `%` comment (`<!-- … -->` in markdown). The part after `--` is the reason, and it stays in the
+file for the next reader:
+
+```latex
+% eslint-disable-next-line paper/leading-zero -- quoted verbatim from the reviewer
+They wrote ".05" in their report.
+Also here: .06 % eslint-disable-line paper/leading-zero -- quoted too
+
+% eslint-disable paper/section-word -- a table reproduced from the venue's own template
+…
+% eslint-enable paper/section-word
+```
+
+This works inside the `filecontents` bibliography of `paper.tex` too, above the entry:
+
+```latex
+% eslint-disable-next-line bib/reachable-entry -- an invited talk, no recording exists
+@misc{smith2024talk, …}
+```
+
+A directive that no longer silences anything is itself reported, so a stale exception does not
+linger. There is no per-paper allowance or counter: the typography rules fix themselves with
+`paperlint lint --fix`, and what is left is either fixed or excepted by name, on its line.
+
 Besides these rules, `paperlint lint` reports a paper directory that is missing a required file (by
 default `PIPELINE-STATUS.md`) as an error. Which files are required is configurable — see
 [`configuration.md`](configuration.md#required-files).
@@ -33,7 +67,7 @@ default `PIPELINE-STATUS.md`) as an error. Which files are required is configura
 ## Checks against the venue
 
 A paper says where it is submitted in a `paperlint.json` beside `paper.tex`
-([`configuration.md`](configuration.md#three-levels-of-settings)):
+([`configuration.md`](configuration.md#a-papers-paperlintjson)):
 
 ```json
 { "extends": "paperlint:aisec", "kind": "research" }
@@ -69,12 +103,13 @@ and the split into body and reference pages. They report on the paper's `paper.t
 | the paper                                                                                        | what you get                                                                                                                 |
 | ------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------- |
 | has a `paperlint.json` that extends no preset (`"extends": null` is what `paperlint new` writes) | `pdf/measured` warning naming the file to set — no venue chosen yet                                                          |
-| has no `paperlint.json` at all (a paper made before 2.1.0)                                       | nothing                                                                                                                      |
+| has no `paperlint.json` at all                                                                   | nothing                                                                                                                      |
 | extends a preset that does not resolve (a typo, a missing file)                                  | `pdf/profile` error, listing the shipped presets                                                                             |
 | has not been built (no facts file)                                                               | `pdf/measured` warning — it does not fail the run, because lint often runs where nothing is built (the CI action only lints) |
 | was built without banal                                                                          | `pdf/measured` warning; fonts are still checked, the rest is not                                                             |
 | has facts about another PDF than the one on disk                                                 | `pdf/fresh` error, and nothing else is judged                                                                                |
-| names no `kind`, or a kind the preset lacks                                                      | `pdf/profile` error; everything but the page limit is still checked                                                          |
+| names no `kind` while its preset has kinds, or a kind the preset lacks                           | `pdf/profile` error; everything but the page limit is still checked                                                          |
+| names no `kind` and its preset has none (`paperlint:acm-sigconf`)                                | nothing — that preset has no page limit to check; everything else is                                                         |
 
 The venue comes from `paperlint.json`, not from the facts, so changing it needs no rebuild: the
 measurements do not depend on it. Messages name the venue by the preset's `name`, else by the file
@@ -100,6 +135,7 @@ Put it in your repository and extend it by a path relative to the file that name
 
 ```
 package.json
+paperlint.json            (optional)
 venues/
   usenix-sec.jsonc
 papers/
@@ -180,8 +216,9 @@ researchQuestion: "Does pruning the state space reduce review cost?"
 ---
 ```
 
-Today you write `bytes:` and `sourceBytes:` yourself, or the agent does, once, when the stage is
-recorded; no command in this package generates them. They are not a checksum to maintain: a
+A stage name may use `a-z`, `0-9`, `.`, `_` and `-`. You write `bytes:` and `sourceBytes:` yourself
+(`wc -c < versions/2026-07-22-submitted.pdf`), or the agent does, once, when the stage is recorded;
+no command in this package generates them. They are not a checksum to maintain: a
 frozen PDF is never supposed to change, so `bytes:` disagreeing with the file means the file was
 replaced after it was declared, and that is exactly the finding. If you really did re-freeze a
 stage, update the number in the same commit.

package/eslint-rules/bib-reachable-entry.mjs

@@ -0,0 +1,75 @@
+/**
+ * `bib/reachable-entry` — a bibliography entry a reader cannot follow: no doi, no url, no arXiv
+ * id. Reported ON THE ENTRY, one finding per entry; there is no fix, because the link has to be
+ * looked up. Default severity: warn.
+ *
+ * Provenance: Reviewer A on HotCRP #20 (2026-08-23), "not every entry has a DOI/link" — one
+ * paper had 0 of 47 entries reachable. Until 3.0.0 this was one count inside `paper/typography`,
+ * silenced per paper by a declared debt; now each entry is its own finding, and an entry that
+ * genuinely has no link (a talk, a personal communication) carries a disable directive with its
+ * reason, in the bibliography itself:
+ *
+ *   % eslint-disable-next-line bib/reachable-entry -- an invited talk, no recording exists
+ *   @misc{smith2024talk, …}
+ *
+ * ⚠️ NOT "no doi". Measured 2026-08-24: ICLR/NeurIPS/TMLR issue no DOIs at all, so a doi-only
+ * rule would demand something that does not exist and get muted for lying. A url or an arXiv id
+ * counts.
+ *
+ * The bibliography read is the one INSIDE `paper.tex` (`filecontents`), where the papers this
+ * package targets keep it. The entry boundaries and fields are found by a line-anchored `@` and
+ * a field name followed by `=` — BibTeX's own lexical shape, one lexeme at a time, not a parse:
+ * the parser (`@retorquere/bibtex-parser`) is an OPTIONAL peer of this package.
+ */
+import { bibRange } from "./paper-typography.mjs";
+
+/** Entry types that are not references. */
+const NOT_AN_ENTRY = /^@(comment|string|preamble)\b/i;
+const HAS_LINK = /\b(doi|url)\s*=/i;
+const HAS_ARXIV = /arxiv[:\s]*\d{4}\.\d{4,5}/i;
+
+export default {
+  rules: {
+    "reachable-entry": {
+      meta: {
+        type: "suggestion",
+        docs: {
+          description:
+            "a bibliography entry carries a doi, a url or an arXiv id — something a reader can follow",
+        },
+        schema: [],
+        messages: {
+          unreachable:
+            "`{{key}}` has no doi, url or arXiv id — a reader has nothing to follow. If none exists, keep the exception with `% eslint-disable-next-line bib/reachable-entry -- <why>` above the entry",
+        },
+      },
+      create(context) {
+        return {
+          "root:exit"() {
+            const sc = context.sourceCode;
+            const raw = sc.raw ?? sc.text;
+            const bib = typeof raw === "string" ? bibRange(raw) : null;
+            if (!bib) return;
+            const starts = [...bib.body.matchAll(/^@/gm)].map((m) => m.index);
+            starts.forEach((s, i) => {
+              const entry = bib.body.slice(s, starts[i + 1] ?? bib.body.length);
+              if (NOT_AN_ENTRY.test(entry)) return;
+              if (HAS_LINK.test(entry) || HAS_ARXIV.test(entry)) return;
+              const head = entry.split("\n")[0];
+              const key = /\{\s*([^,\s]+)/.exec(head)?.[1] ?? "?";
+              const from = bib.bodyStart + s;
+              context.report({
+                loc: {
+                  start: sc.getLocFromIndex(from),
+                  end: sc.getLocFromIndex(from + head.length),
+                },
+                messageId: "unreachable",
+                data: { key },
+              });
+            });
+          },
+        };
+      },
+    },
+  },
+};

package/eslint-rules/latex-language.mjs

@@ -64,7 +64,9 @@
  *     is not substituted). Anonymisation is normally done with exactly those, so coinage and
  *     jargon rules see less on an anonymised version than on a camera-ready.
  *  8. `% eslint {"rules":…}` inside `.tex` does NOT work: `applyInlineConfig` returns empty.
- *     `% eslint-disable-next-line <rule> -- reason` DOES work.
+ *     The disable directives DO work — `% eslint-disable-next-line <rule> -- reason`,
+ *     `% eslint-disable-line`, `% eslint-disable` / `% eslint-enable` — and since 3.0.0 also
+ *     inside a `filecontents` bibliography, whose `%` lines are made comment nodes below.
  *  9. `\section*{…}` is untested and does not occur in the corpus (measured: 0 occurrences).
  *     When it appears, `plain()` will glue all arguments together and the heading may come out
  *     as `*X`.
@@ -207,6 +209,32 @@ export function texToMdast(src) {
     });
   }
 
+  // 🔴 A `filecontents` body is not LaTeX — it is the `.bib` the paper writes, usually in the
+  // preamble — so the parser returns it as one verbatim node and no comment nodes from it. A
+  // disable directive above a bibliography entry (`% eslint-disable-next-line bib/reachable-entry
+  // -- …`) was therefore invisible. Its `%` lines are made comments here, the only way
+  // `getInlineConfigNodes` sees them. BibTeX ignores text between entries, so the line is
+  // harmless to the build.
+  for (const n of ast.content)
+    if (
+      n.type === "verbatim" &&
+      /^filecontents\*?$/.test(String(n.env)) &&
+      P(n)
+    ) {
+      let at = P(n).start.offset;
+      for (const line of src.slice(at, P(n).end.offset).split("\n")) {
+        const text = line.replace(/\r$/, "");
+        const pct = /^\s*%/.test(text) ? text.indexOf("%") : -1;
+        if (pct >= 0)
+          children.push({
+            type: "html",
+            value: text.slice(pct + 1),
+            position: { start: loc(at + pct), end: loc(at + text.length) },
+          });
+        at += line.length + 1;
+      }
+    }
+
   (function walk(node) {
     if (Array.isArray(node)) return node.forEach((n) => walk(n));
     if (!node || typeof node !== "object") return;

package/eslint-rules/paper-stages.harness.mjs

@@ -45,7 +45,7 @@ const since = () => {
 
 assert.deepEqual(
   Object.keys(stages.rules).sort(),
-  ["author-list", "source", "stages"],
+  ["source", "stages"],
   "rule set changed",
 );
 
@@ -237,120 +237,4 @@ console.log(
   `✓ ${String(since())} assertions passed — paper/source, frozen bytes instead of a sha`,
 );
 
-// ── `paper/author-list`: a shipped paper owes itself a run of the author-list cross-check ────
-//
-// This rule's subject is DIFFERENT from its two neighbors above: those check the declaration
-// against the bytes, this one checks the declaration against a record of a run. What they share
-// is exactly one input, the `stages` field, and that's why the rule lives in this module.
-{
-  const lintAuthors = (name, opts = {}) => {
-    const file = join(FIX, name, "PIPELINE-STATUS.md");
-    const msgs = linter.verify(
-      readFileSync(file, "utf-8"),
-      [
-        {
-          files: ["**/*.md"],
-          plugins: { markdown, paper: stages },
-          language: "markdown/gfm",
-          languageOptions: { frontmatter: "yaml" },
-          rules: { "paper/author-list": ["error", opts] },
-        },
-      ],
-      file,
-    );
-    assert.deepEqual(
-      msgs.filter((m) => m.fatal),
-      [],
-      `${name}: the rule threw`,
-    );
-    return msgs.map((m) => m.message);
-  };
-
-  // ── stays silent where it must ──
-  check(
-    "a run is recorded in the scorecard — silent",
-    lintAuthors("authors-ran").length === 0,
-  );
-  // A draft never asked anyone to read it, so it owes nothing. This is not a carve-out: the
-  // rule's subject is the DEBT of a shipped paper, and an unshipped one has no debt.
-  check(
-    "no stage declared at all — silent, a draft owes nothing",
-    lintAuthors("nothing").length === 0,
-  );
-  // An empty list is not "a stage exists": the record `stages: []` shows up on a paper that was
-  // set up but never submitted anywhere.
-  check(
-    "an empty stage list — silent",
-    linter.verify(
-      "---\nstages: []\n---\n# S\n",
-      [
-        {
-          files: ["**/*.md"],
-          plugins: { markdown, paper: stages },
-          language: "markdown/gfm",
-          languageOptions: { frontmatter: "yaml" },
-          rules: { "paper/author-list": "error" },
-        },
-      ],
-      join(FIX, "x", "PIPELINE-STATUS.md"),
-    ).length === 0,
-  );
-
-  // ── fires on a planted defect ──
-  const owed = lintAuthors("ok");
-  check("a stage is declared, no run — a finding", owed.length === 1);
-  // 🔴 The message must name the CLASS, not just the fact of a miss: otherwise the reader takes
-  // it for a duplicate of the citation-existence check and closes it as noise. The class is
-  // preprint authors under a declared conference, and it is invisible to a check that a
-  // citation merely resolves.
-  check(
-    "and it names the class the cross-check catches, not just the miss",
-    /PREPRINT/.test(owed[0]),
-  );
-
-  // ── the whole reason the move was made ──
-  // The predecessor derived the stage with a REGEX OVER THE SCORECARD'S PROSE. Remeasured
-  // 09-17: for `agenticdev-2026` the prose reads `submitted`, while the frontmatter reads
-  // `submitted, camera-ready`. Here the list comes from the field, so both stages land in the
-  // finding's text.
-  const two = lintAuthors("twice");
-  check(
-    "the stage list in the message comes from the FIELD and carries all of them",
-    two.length === 1 && /submitted\/submitted/.test(two[0]),
-  );
-
-  // ── consumer data stays with the consumer ──
-  // The predecessor hardcoded the path `.claude/skills/verify-citations/...` into the message
-  // text — the address of ONE repository inside a public package.
-  const withCmd = lintAuthors("ok", {
-    command: "node scripts/bib-authors.mjs <paper>",
-  });
-  check(
-    "the run command comes in as an option and lands in the message",
-    /scripts\/bib-authors\.mjs/.test(withCmd[0]),
-  );
-  check(
-    "and without the option the message does not invent a path",
-    !/bib-authors\.mjs/.test(owed[0]),
-  );
-  // The marker is also data: the package cannot know EXACTLY how a given consumer records a run.
-  check(
-    "the marker is configurable — with a different marker the same paper becomes a debtor",
-    lintAuthors("authors-ran", { marker: "no-such-marker" }).length === 1,
-  );
-
-  // ── parsing versus grep: the one case where they diverge ──
-  // 🔴 This assert is the proof of the move to a parser. All the fixtures above pass IDENTICALLY
-  // either way, because their marker sits in a cell. Here it sits in PROSE — "still need to run
-  // bib-authors", an intention, not a record — and grep would read it as evidence of a run. The
-  // evidence must sit in the scorecard.
-  check(
-    "a marker in prose OUTSIDE the table is not a record of a run",
-    lintAuthors("marker-in-prose").length === 1,
-  );
-}
-console.log(
-  `✓ ${String(since())} assertions passed — paper/author-list, the debt of a shipped paper`,
-);
-
 console.log(`✓ ${String(n)} assertions passed in total`);

package/eslint-rules/paper-stages.mjs

@@ -254,126 +254,6 @@ export default {
      * because a stage frozen BEFORE this convention existed cannot be fixed at all — those
      * bytes are gone. Failing a build over unrecoverable history is a gate nobody can clear.
      */
-    /**
-     * `paper/author-list` — a shipped paper OWES ITSELF a run of the author-list cross-check.
-     *
-     * The class this cross-check catches is invisible to an existence check on citations: the
-     * citation exists, the identifier resolves, and the authors are taken from the PREPRINT
-     * while the entry declares a conference. On the live corpus this turned up seven records
-     * across three papers, including a DROPPED LIVING PERSON (`schick2023toolformer` — Eric
-     * Hambro is missing; the NeurIPS version has nine authors, the preprint has eight). Two of
-     * the seven were found on a paper ALREADY SUBMITTED.
-     *
-     * 🔴 WHAT THE MOVE FIXED, and this is a measurement, not taste. The predecessor derived the
-     * declared stage with a REGEX OVER THE PROSE of that same file. Remeasured 2026-09-17 on the
-     * live corpus: for `agenticdev-2026` the prose reads `submitted`, the frontmatter declares
-     * `submitted, camera-ready` — the pattern `/camera-ready (?:uploaded|submitted|shipped)/i`
-     * does not catch the form the stage is actually recorded in. Today's set of findings did not
-     * change from this (both checks only asked "is there ANY stage at all"), but the MESSAGE
-     * printed the wrong list of stages. The subject here is the same `stages` field that
-     * `paper/stages` already checks against the bytes in both directions.
-     *
-     * 🔴 THE MARKER IS LOOKED FOR IN TABLE CELLS, NOT GREPPED OVER THE FILE. The first draft did
-     * `context.sourceCode.text.includes(marker)` and justified it with a comment saying "a
-     * footnote cell has no node of its own". That turned out to be SIMPLY WRONG — a 2026-09-17
-     * measurement showed the markdown parser hands back `table`, `tableRow` and `tableCell`
-     * (twelve cells on a three-row table). The base rule says, word for word, "parse markdown
-     * with a parser".
-     *
-     * Parsing is also STRICTER than grep, and the difference is real: a marker mentioned in
-     * prose outside the scorecard — in a heading, in a paragraph saying "still need to run
-     * bib-authors", in a quote from someone else — no longer counts as a record of a run. To
-     * grep, those three cases are indistinguishable from a genuine record.
-     *
-     * ⚠️ WHAT THIS STILL DOES NOT FIX: inside the cell the evidence is still PROSE, and a run
-     * phrased in different words will not be seen by the rule. The real fix is a frontmatter
-     * field (`gates.cites.ran`), not a smarter text search; that's separate work, touching four
-     * live scorecards.
-     *
-     * So severity is set by the CONSUMER, and the default is not `error`: the proof of a run is
-     * prose, and a false positive at a blocking level costs more than a miss.
-     */
-    "author-list": {
-      meta: {
-        type: "suggestion",
-        docs: {
-          description:
-            "a paper that declares a stage records that the author-list check ran — a class the existence check cannot see",
-        },
-        schema: [
-          {
-            type: "object",
-            properties: {
-              // The run marker in the scorecard. This is consumer data: the package cannot
-              // know EXACTLY how a given consumer records that the cross-check happened.
-              marker: { type: "string" },
-              // What to run. This is a CONSUMER-SPECIFIC PATH, and it has no place in a public
-              // package: the predecessor hardcoded
-              // `.claude/skills/verify-citations/scripts/bib-authors.mjs` right into the message text.
-              command: { type: "string" },
-            },
-            additionalProperties: false,
-          },
-        ],
-        messages: {
-          neverRan:
-            "stage «{{stages}}» is declared, but the scorecard records no author-list run (looked for «{{marker}}» in its table). It catches what an existence check cannot see: the citation resolves, the id resolves, and the authors are the PREPRINT's while the entry declares a conference{{how}}",
-        },
-      },
-      create(context) {
-        const opts = context.options?.[0] ?? {};
-        const marker = opts.marker ?? "bib-authors";
-        const command = opts.command ?? "";
-
-        // The verdict is deferred to the end of the file: the frontmatter node arrives FIRST,
-        // and the table comes after it. Reporting on `yaml` would mean a verdict rendered
-        // without ever seeing the scorecard.
-        let declaredAt = null;
-        let stages = [];
-        let recorded = false;
-
-        return {
-          yaml(node) {
-            let data;
-            try {
-              data = load(node.value ?? "");
-            } catch {
-              return; // `paper/stages` has already reported the unreadable YAML
-            }
-            const raw = data?.stages;
-            if (!Array.isArray(raw)) return;
-            stages = raw.map((r) => r?.stage).filter(Boolean);
-            // One guard, not two: `raw.length === 0` would be a SPECIAL CASE of this same
-            // condition, and a mutation against it would turn out unkillable — the second guard
-            // mutes it. This also covers a non-empty list made of records with no `stage` field.
-            if (stages.length === 0) return; // nothing shipped — nothing is owed
-            declaredAt = node;
-          },
-
-          // The evidence is a SCORECARD CELL, not any occurrence of the string in the file. It
-          // has a node of its own; the first draft claimed otherwise and grepped the whole text.
-          tableCell(node) {
-            if (recorded) return;
-            if (context.sourceCode.getText(node).includes(marker))
-              recorded = true;
-          },
-
-          "root:exit"() {
-            if (!declaredAt || recorded) return;
-            context.report({
-              node: declaredAt,
-              messageId: "neverRan",
-              data: {
-                stages: stages.join("/"),
-                marker,
-                how: command ? `. Run: ${command}` : "",
-              },
-            });
-          },
-        };
-      },
-    },
-
     source: {
       meta: {
         type: "problem",

package/eslint-rules/paper-stages.mutations.mjs

@@ -76,82 +76,6 @@ process.exit(
           ],
         ],
       },
-      {
-        // 🔴 A REGRESSION BACK TO GREP. The rule's first draft did exactly this and justified it
-        // by saying "a footnote cell has no node of its own" — a measurement showed the
-        // opposite: the parser hands back `tableCell`. The mutation is caught by ONE fixture out
-        // of six: in the rest the marker already sits in a cell, so grep and parsing are
-        // indistinguishable. Without `marker-in-prose` this regression would have passed
-        // silently, and the rule would again read an intention as a fact.
-        name: "the evidence is grepped over the WHOLE FILE again",
-        harness: HARNESS,
-        expect: "a marker in prose OUTSIDE the table is not a record of a run",
-        disables:
-          "parsing in favor of a string search: 'still need to run bib-authors' — an intention, " +
-          "not a record — counts as a run again, and the paper silently stops being a debtor",
-        edits: [
-          [
-            RULE,
-            "if (context.sourceCode.getText(node).includes(marker))\n              recorded = true;",
-            "if (context.sourceCode.text.includes(marker))\n              recorded = true;",
-          ],
-        ],
-      },
-      {
-        name: "the rule stops requiring an author cross-check run",
-        harness: HARNESS,
-        expect: "a stage is declared, no run — a finding",
-        disables:
-          "the debt itself: a shipped paper no longer owes anything, and the rule stays silent " +
-          "across the whole corpus — silence is its success state, so from the outside this is " +
-          "indistinguishable",
-        // Target retargeted 09-17: the old line disappeared when the rule moved to parsing, and
-        // the driver honestly said UNUSABLE — "the patch did not land", not "the test is weak".
-        // Now the mutation declares the run recorded before the file is even read.
-        edits: [[RULE, "let recorded = false;", "let recorded = true;"]],
-      },
-      {
-        name: "the exemption for an unshipped paper becomes WIDER than it should",
-        harness: HARNESS,
-        expect: "an empty stage list — silent",
-        disables:
-          "the distinction between 'no stages' and 'there are stages': a draft starts getting a " +
-          "finding, and a rule that scolds drafts gets turned off within a week",
-        edits: [
-          [
-            RULE,
-            "if (stages.length === 0) return; // nothing shipped — nothing is owed",
-            "",
-          ],
-        ],
-      },
-      {
-        name: "the stage list is taken from somewhere OTHER than the field again",
-        harness: HARNESS,
-        expect:
-          "the stage list in the message comes from the FIELD and carries all of them",
-        disables:
-          "the exact reason the move was made. The predecessor derived the stage with a regex " +
-          "over the prose, and on agenticdev printed `submitted` where `submitted, " +
-          "camera-ready` was declared. The mutation brings back a hardcoded list — the set of " +
-          "findings does not change, only the TEXT lies, and without this assert the regression " +
-          "would have passed silently",
-        edits: [[RULE, 'stages: stages.join("/"),', 'stages: "submitted",']],
-      },
-      {
-        name: "a consumer-specific path comes back into the package's text",
-        harness: HARNESS,
-        expect:
-          "the run command comes in as an option and lands in the message",
-        disables:
-          "the boundary 'the mechanism goes in the package, the data stays with the consumer': " +
-          "the command stops arriving as an option. The predecessor hardcoded " +
-          "`.claude/skills/verify-citations/...` — the path of one private repository — right " +
-          "into a public rule's message",
-        edits: [
-          [RULE, 'const command = opts.command ?? "";', 'const command = "";'],
-        ],
-      },
     ],
   }),
 );