paperlint 7.2.0 → 8.0.0

package/skills/tighten-paper/SKILL.md

@@ -4,7 +4,7 @@ description: "Use when a paper \"feels long / bloated / overly complex / hard to
 allowed-tools: [Read, Write, Edit, Grep, Glob, Agent]
 ---
 
-<!-- vigiles:sha256:fd00cde33ea4b81c compiled from skills/tighten-paper/SKILL.md.spec.ts -->
+<!-- vigiles:sha256:c57d6da372b28acf compiled from skills/tighten-paper/SKILL.md.spec.ts -->
 
 # tighten-paper — the developmental edit
 
@@ -34,31 +34,12 @@ counter, is not this skill; it is the damage this skill exists to undo.
 If you are invoked while the author is still adding material, say so and decline the pass. The right
 answer is "not yet", not a smaller cut.
 
-## 🏗 START WITH THE MECHANICAL LEG — run `structure.mjs` (added 2026-08-05)
+## 🏗 START WITH THE SHAPE — weigh every section before judging it
 
-```
-node .claude/skills/tighten-paper/structure.mjs <paper.md>              # whole paper
-node .claude/skills/tighten-paper/structure.mjs <paper.md> --section=3  # one section
-```
-
-🔴 **THE SCRIPT NO LONGER FINDS ANYTHING — it is an INVENTORY (2026-08-26).** All sixteen of its
-checks moved into ESLint rules (`eslint-rules/paper-structure.mjs`). The findings you used to read
-here now come from:
-
-```
-npx eslint --no-config-lookup --config eslint.config.mjs <paper.md>
-```
-
-⚠️ Do NOT pass `--flags-only`. It is kept only because two harnesses assert on its silence, and it
-prints nothing and exits 0 whatever the paper looks like. Measured 26.08 on `the reference paper`:
-`--flags-only` printed 0 lines while ESLint on the same bytes reported **6 structural findings**
-(subsection-size ×2 · section-lead · free-section-size ×2 · block-ungraded). A mode that cannot
-speak is not a check, and reading its silence as "clean" is the defect this pipeline is named for.
-
-What the inventory prints, and why it stays a script: the outline **in order** — every heading, its
-own words, its total with subsections, its share of the body, and the `carries:`/`score:`/`verdict:`
-note the section wrote about itself, printed beside its weight. That comparison is the whole point
-and a linter cannot make it: a finding has nowhere to put a table.
+Write the outline out **in order**: every heading, its own words before its first subsection, its
+total with subsections, its share of the body — and beside each, the `carries:`/`score:`/`verdict:`
+note the section wrote about itself. That comparison, a section's weight printed next to what it
+claims about itself, is the whole point.
 
 🔴 **Why this exists, and why word counts alone mislead.** Asked whether a paper short-changed its
 contribution, per-section word counts said no — the contribution had 38% more words than related
@@ -67,7 +48,7 @@ owned **0, 28 and 38** words before their first subsection. Nothing in the toolc
 question, because every metric was per-sentence or per-document and none was per-section. The author's
 word for the result was a "salad" — a jumbled mix — and he was right while the numbers said fine.
 
-**What the script does NOT decide, and you must:** the ORDER, the NAMES, and whether a section earns
+**What the count does NOT decide, and you must:** the ORDER, the NAMES, and whether a section earns
 its place. Those are judgement, and pretending otherwise would be the failure this whole pipeline
 exists to prevent.
 
@@ -188,18 +169,24 @@ called it "fat for no reason" after a pass that cut almost nothing and left it 1
 
 A verdict that lives in a chat log is gone by the next session, and the next session then re-derives
 it or ignores it. **Every section and subsection gets a one-line verdict comment immediately above its
-heading**, in the paper source, stamped with the date and the skill that produced it:
-
-```
-<!-- TIGHTEN 2026-08-06 · score 8/10 · verdict: KEEP · 190w
-     carries: tools get close but none resolves a NAMED rule against a committed config
-     because: it is the only place the tool-ecosystem gap is closed; a reviewer who skips it
-              asks "hasn't this been done" and gets no answer -->
-<!-- TIGHTEN 2026-08-06 · score 5/10 · verdict: SHORTEN -89w · 210w
-     carries: why a linter for the rules file cannot work
-     because: the point lands in two sentences; the rest re-establishes what §2 already did -->
-<!-- TIGHTEN 2026-08-06 · score 2/10 · verdict: MOVE→artifact · 45w
-     carries: nothing the reader must hold to judge the contribution -->
+heading**, in `paper.tex`, as LaTeX `%` comments, stamped with the date and the skill that
+produced it:
+
+```latex
+% TIGHTEN 2026-08-06 · score 8/10 · verdict: KEEP · 190w
+%   carries: tools get close but none resolves a NAMED rule against a committed config
+%   because: it is the only place the tool-ecosystem gap is closed; a reviewer who skips it
+%            asks "hasn't this been done" and gets no answer
+\section{Related tools}
+
+% TIGHTEN 2026-08-06 · score 5/10 · verdict: SHORTEN -89w · 210w
+%   carries: why a linter for the rules file cannot work
+%   because: the point lands in two sentences; the rest re-establishes what §2 already did
+\subsection{Why not a linter}
+
+% TIGHTEN 2026-08-06 · score 2/10 · verdict: MOVE→artifact · 45w
+%   carries: nothing the reader must hold to judge the contribution
+\subsection{Implementation notes}
 ```
 
 🔴 **Before assigning any `MOVE→` verdict, read `paper-pipeline/references/body-vs-appendix.md`.**
@@ -224,7 +211,7 @@ written and **not one ever concluded "cut me"**. Author, 2026-08-06, reading the
 value on every page and compared nothing.
 
 The pairing is the mechanism: **a low score beside a KEEP verdict is a visible contradiction**, and
-`structure.mjs` reports it. Prose cannot contradict itself; a number can.
+the outline above puts the two side by side. Prose cannot contradict itself; a number can.
 
 **Calibration, so the scale does not drift to "everything is an 8":**
 - **9–10** — remove it and a reviewer cannot assess the contribution. The claim, its evidence, its price.
@@ -279,10 +266,10 @@ Rules for these comments:
   unpriced prose accumulates, because nothing there costs a page.
 - **Stale beats absent, but not by much.** A comment whose date precedes the last substantive edit to
   its section is stale; re-grade it rather than trusting it.
-- **A verdict nobody executed is a note, not a decision.** `structure.mjs` reports blocks still
-  carrying a non-KEEP verdict — a paper full of unexecuted CUTs is worse than one never graded,
+- **A verdict nobody executed is a note, not a decision.** List the blocks still carrying a
+  non-KEEP verdict — a paper full of unexecuted CUTs is worse than one never graded,
   because it reads as decided.
-- They are working notes: stripped before typesetting, like every other `<!-- -->` block.
+- They are working notes: a `%` comment never reaches the PDF.
 - The chat summary still gets written — but it is the *derivative*, and the paper is the record.
 
 ## 🧱 Enforce subsections — a section without them is a wall

package/skills/tighten-paper/SKILL.md.spec.ts

@@ -40,31 +40,12 @@ counter, is not this skill; it is the damage this skill exists to undo.
 If you are invoked while the author is still adding material, say so and decline the pass. The right
 answer is "not yet", not a smaller cut.
 
-## 🏗 START WITH THE MECHANICAL LEG — run \`structure.mjs\` (added 2026-08-05)
+## 🏗 START WITH THE SHAPE — weigh every section before judging it
 
-\`\`\`
-node .claude/skills/tighten-paper/structure.mjs <paper.md>              # whole paper
-node .claude/skills/tighten-paper/structure.mjs <paper.md> --section=3  # one section
-\`\`\`
-
-🔴 **THE SCRIPT NO LONGER FINDS ANYTHING — it is an INVENTORY (2026-08-26).** All sixteen of its
-checks moved into ESLint rules (\`eslint-rules/paper-structure.mjs\`). The findings you used to read
-here now come from:
-
-\`\`\`
-npx eslint --no-config-lookup --config eslint.config.mjs <paper.md>
-\`\`\`
-
-⚠️ Do NOT pass \`--flags-only\`. It is kept only because two harnesses assert on its silence, and it
-prints nothing and exits 0 whatever the paper looks like. Measured 26.08 on \`the reference paper\`:
-\`--flags-only\` printed 0 lines while ESLint on the same bytes reported **6 structural findings**
-(subsection-size ×2 · section-lead · free-section-size ×2 · block-ungraded). A mode that cannot
-speak is not a check, and reading its silence as "clean" is the defect this pipeline is named for.
-
-What the inventory prints, and why it stays a script: the outline **in order** — every heading, its
-own words, its total with subsections, its share of the body, and the \`carries:\`/\`score:\`/\`verdict:\`
-note the section wrote about itself, printed beside its weight. That comparison is the whole point
-and a linter cannot make it: a finding has nowhere to put a table.
+Write the outline out **in order**: every heading, its own words before its first subsection, its
+total with subsections, its share of the body — and beside each, the \`carries:\`/\`score:\`/\`verdict:\`
+note the section wrote about itself. That comparison, a section's weight printed next to what it
+claims about itself, is the whole point.
 
 🔴 **Why this exists, and why word counts alone mislead.** Asked whether a paper short-changed its
 contribution, per-section word counts said no — the contribution had 38% more words than related
@@ -73,7 +54,7 @@ owned **0, 28 and 38** words before their first subsection. Nothing in the toolc
 question, because every metric was per-sentence or per-document and none was per-section. The author's
 word for the result was a "salad" — a jumbled mix — and he was right while the numbers said fine.
 
-**What the script does NOT decide, and you must:** the ORDER, the NAMES, and whether a section earns
+**What the count does NOT decide, and you must:** the ORDER, the NAMES, and whether a section earns
 its place. Those are judgement, and pretending otherwise would be the failure this whole pipeline
 exists to prevent.
 
@@ -194,18 +175,24 @@ called it "fat for no reason" after a pass that cut almost nothing and left it 1
 
 A verdict that lives in a chat log is gone by the next session, and the next session then re-derives
 it or ignores it. **Every section and subsection gets a one-line verdict comment immediately above its
-heading**, in the paper source, stamped with the date and the skill that produced it:
-
-\`\`\`
-<!-- TIGHTEN 2026-08-06 · score 8/10 · verdict: KEEP · 190w
-     carries: tools get close but none resolves a NAMED rule against a committed config
-     because: it is the only place the tool-ecosystem gap is closed; a reviewer who skips it
-              asks "hasn't this been done" and gets no answer -->
-<!-- TIGHTEN 2026-08-06 · score 5/10 · verdict: SHORTEN -89w · 210w
-     carries: why a linter for the rules file cannot work
-     because: the point lands in two sentences; the rest re-establishes what §2 already did -->
-<!-- TIGHTEN 2026-08-06 · score 2/10 · verdict: MOVE→artifact · 45w
-     carries: nothing the reader must hold to judge the contribution -->
+heading**, in \`paper.tex\`, as LaTeX \`%\` comments, stamped with the date and the skill that
+produced it:
+
+\`\`\`latex
+% TIGHTEN 2026-08-06 · score 8/10 · verdict: KEEP · 190w
+%   carries: tools get close but none resolves a NAMED rule against a committed config
+%   because: it is the only place the tool-ecosystem gap is closed; a reviewer who skips it
+%            asks "hasn't this been done" and gets no answer
+\\section{Related tools}
+
+% TIGHTEN 2026-08-06 · score 5/10 · verdict: SHORTEN -89w · 210w
+%   carries: why a linter for the rules file cannot work
+%   because: the point lands in two sentences; the rest re-establishes what §2 already did
+\\subsection{Why not a linter}
+
+% TIGHTEN 2026-08-06 · score 2/10 · verdict: MOVE→artifact · 45w
+%   carries: nothing the reader must hold to judge the contribution
+\\subsection{Implementation notes}
 \`\`\`
 
 🔴 **Before assigning any \`MOVE→\` verdict, read \`paper-pipeline/references/body-vs-appendix.md\`.**
@@ -230,7 +217,7 @@ written and **not one ever concluded "cut me"**. Author, 2026-08-06, reading the
 value on every page and compared nothing.
 
 The pairing is the mechanism: **a low score beside a KEEP verdict is a visible contradiction**, and
-\`structure.mjs\` reports it. Prose cannot contradict itself; a number can.
+the outline above puts the two side by side. Prose cannot contradict itself; a number can.
 
 **Calibration, so the scale does not drift to "everything is an 8":**
 - **9–10** — remove it and a reviewer cannot assess the contribution. The claim, its evidence, its price.
@@ -285,10 +272,10 @@ Rules for these comments:
   unpriced prose accumulates, because nothing there costs a page.
 - **Stale beats absent, but not by much.** A comment whose date precedes the last substantive edit to
   its section is stale; re-grade it rather than trusting it.
-- **A verdict nobody executed is a note, not a decision.** \`structure.mjs\` reports blocks still
-  carrying a non-KEEP verdict — a paper full of unexecuted CUTs is worse than one never graded,
+- **A verdict nobody executed is a note, not a decision.** List the blocks still carrying a
+  non-KEEP verdict — a paper full of unexecuted CUTs is worse than one never graded,
   because it reads as decided.
-- They are working notes: stripped before typesetting, like every other \`<!-- -->\` block.
+- They are working notes: a \`%\` comment never reaches the PDF.
 - The chat summary still gets written — but it is the *derivative*, and the paper is the record.
 
 ## 🧱 Enforce subsections — a section without them is a wall

package/skills/verify-citations/scripts/bib-authors.check.test.mjs

@@ -50,15 +50,6 @@ test("checkAuthors sorts every kind of entry into its bucket", async () => {
     entry("nomatch", "Ada Lovelace", "Nobody Has This"),
     entry("onlycorr", "Ada Lovelace", "Only On CoRR"),
     entry("down", "Ada Lovelace", "Service Down"),
-    {
-      type: "mdref",
-      key: "sw",
-      unparsed: true,
-      author: "",
-      title: "",
-      booktitle: "",
-      journal: "",
-    },
   ];
   const answers = {
     "Same Paper": [hit("Same Paper.", ["Ada Lovelace", "Alan Turing"])],
@@ -101,10 +92,6 @@ test("checkAuthors sorts every kind of entry into its bucket", async () => {
         },
       ],
       skipped: [
-        {
-          key: "sw",
-          why: "no author/title — a reference to software or a dataset, DBLP does not apply",
-        },
         {
           key: "etal",
           why: "author list ends in `and others` — completeness not checkable",

package/skills/verify-citations/scripts/bib-authors.cli.test.ts

@@ -1,7 +1,7 @@
 /**
- * bib-authors.mjs as a process: which file it picks from a paper directory, the three
- * bibliography sources it reads (.bib, a .tex's filecontents or its sibling refs.bib, a markdown
- * reference list), and the report and exit code for each outcome. DBLP is a preload that answers
+ * bib-authors.mjs as a process: which file it picks from a paper directory, the bibliography
+ * sources it reads (.bib, a .tex's filecontents or its sibling refs.bib), the file it refuses, and
+ * the report and exit code for each outcome. DBLP is a preload that answers
  * by title from a table (the documented `result.hits.hit[].info` shape) — no network.
  */
 import assert from "node:assert/strict";
@@ -58,8 +58,10 @@ writeTree(root, {
   "sibling/paper.tex": "\\documentclass{article}\n",
   "sibling/refs.bib": bib("good", "Ada Lovelace", "All Good"),
   "nobib/paper.tex": "\\documentclass{article}\n",
-  "md/paper.md":
-    "# P\n\n## References\n\n1. A. Lovelace. *All Good.* Proc. X, 2024.\n2. some-npm-package, https://npm.example\n",
+  "etal/notes.txt": "not a bibliography\n",
+  "etal/refs.bib":
+    bib("good", "Ada Lovelace", "All Good") +
+    bib("etal", "Ada Lovelace and others", "Truncated"),
   "down/refs.bib": bib("down", "Ada Lovelace", "Service Down"),
   "empty/.keep": "",
 });
@@ -98,7 +100,7 @@ test("refusals: no argument, no such path, a directory with nothing to read, a .
       {
         status: 2,
         stdout: "",
-        stderr: "bib-authors: no .bib, .tex or paper.md in <root>/empty\n",
+        stderr: "bib-authors: no .bib or .tex in <root>/empty\n",
       },
       {
         status: 2,
@@ -162,15 +164,25 @@ test("a .tex with no filecontents reads the refs.bib beside it — PASS, exit 0"
   });
 });
 
-test("a markdown reference list: a software entry is not applicable, not dropped", () => {
-  const r = run(join(root, "md"));
+test("a file that is neither a .bib nor a .tex is refused — not swapped for the refs.bib beside it", () => {
+  const r = run(join(root, "etal", "notes.txt"));
+  assert.deepEqual(r, {
+    status: 2,
+    stdout: "",
+    stderr:
+      "bib-authors: <root>/etal/notes.txt is not a .bib or a .tex — give the paper's .bib or paper.tex\n",
+  });
+});
+
+test("an `and others` entry is printed as not applicable, not dropped", () => {
+  const r = run(join(root, "etal"));
   assert.deepEqual(r, {
     status: 0,
     stderr: "",
     stdout:
-      "== bib-authors: <root>/md/paper.md ==\n" +
-      "  · n/a ref2 — no author/title — a reference to software or a dataset, DBLP does not apply\n\n" +
-      "-- 1 entries · 0 difference(s) · 1 not applicable · 0 NOT CHECKED\nPASS: no author-list disagreement.\n",
+      "== bib-authors: <root>/etal/refs.bib ==\n" +
+      "  · n/a etal — author list ends in `and others` — completeness not checkable\n\n" +
+      "-- 2 entries · 0 difference(s) · 1 not applicable · 0 NOT CHECKED\nPASS: no author-list disagreement.\n",
   });
 });
 

package/skills/verify-citations/scripts/bib-authors.mjs

@@ -45,90 +45,12 @@
  */
 import { readFileSync, existsSync, readdirSync } from "node:fs";
 import { join, extname } from "node:path";
-import { headings as mdHeadings, requireMarkdown } from "#lib/markdown";
 import { isMain } from "../../paper-pipeline/scripts/consumer.mjs";
 
 const DBLP = "https://dblp.org/search/publ/api";
 
 /* ---------- input: a .bib, a .tex with filecontents, or a paper dir ---------- */
 
-/**
- * Some papers carry no BibTeX at all. `compile-rules-2026` writes its 67 references as a
- * hand-numbered markdown list that `repro/md2acl.py` turns into ACL format at build time:
- *
- *   12. N. F. Liu, K. Lin, J. Hewitt. *Lost in the Middle: How Language Models Use Long
- *       Contexts.* TACL, 2024. arXiv:2307.03172.
- *
- * 🔴 Exempting such a paper would be the wrong call, and measurably so: a hand-written author
- * list has NO machine-checkable source, so it is the MORE exposed of the two formats, not the
- * less. The class this script hunts lives exactly where an entry names a published venue and an
- * arXiv id in the same breath — which is most of this list.
- *
- * Authors are taken only from the text BEFORE the first `*`, on purpose: entries here sometimes
- * append a second work ("— and the reanalysis: S. B. Hossain, ...") after the title, and folding
- * those names in would invent co-authors that the first work does not have.
- */
-export function parseMarkdownRefs(text) {
-  // 🔴 The heading comes FROM THE PARSER, not from a regex (the base's rule "parse markdown with a
-  // parser"). The previous `text.search(/^#+\s*References\s*$/m)` opened the bibliography on a
-  // `## References` line QUOTED inside a ``` fence — and the papers in this repo quote their own
-  // markup in chunks. Exactly this assert already stands at a neighbour
-  // (`paper-pipeline/scripts/extract-ref-facts.harness.mjs`: "a heading inside a ``` block is not a
-  // heading"), that is, the class was known and was being reproduced here again.
-  // `requireMarkdown()` — so that a missing markdown-it fails LOUDLY: with an empty parser
-  // `headings()` returns [], and the gate would report "no references" instead of refusing. That is
-  // the same trade of one silent failure for another that the rule was written against.
-  requireMarkdown();
-  const h = mdHeadings(text).find((x) => /^References\s*$/u.test(x.text));
-  if (!h) return [];
-  const body = text.slice(h.offset);
-  const out = [];
-  // 🔴 `$(?![\s\S])`, NOT `\Z`. JavaScript has no `\Z` anchor — outside a unicode-mode pattern it
-  // is an identity escape meaning the LETTER Z, so the first version of this lookahead ended entry
-  // 1 in the middle of "J. Zhou", at the Z. Measured: 48 of 67 references parsed, and the missing
-  // 19 looked like ordinary gaps rather than a bug. Same family as the `\b`-over-Cyrillic defect
-  // already recorded in this repo: an escape that means one thing in Perl/Python and another here.
-  const re = /^(\d+)\.[ \t]+([\s\S]*?)(?=^\d+\.[ \t]|^#|$(?![\s\S]))/gm;
-  let m;
-  while ((m = re.exec(body))) {
-    const entry = m[2].replace(/\s*\n\s*/g, " ").trim();
-    const t = entry.match(/\*([^*]+)\*/);
-    if (!t) {
-      // No italic title and no author list — in this corpus these are SOFTWARE references
-      // (npm packages, repos). Nothing for DBLP to disagree with. They are surfaced as
-      // not-applicable rather than dropped: an entry that silently vanishes between the file
-      // and the report is indistinguishable from an entry that passed.
-      out.push({
-        type: "mdref",
-        key: `ref${m[1]}`,
-        unparsed: true,
-        author: "",
-        title: "",
-        booktitle: "",
-        journal: "",
-      });
-      continue;
-    }
-    const authorPart = entry
-      .slice(0, entry.indexOf("*"))
-      .trim()
-      .replace(/[.,;]\s*$/, "");
-    const rest = entry.slice(entry.indexOf(t[0]) + t[0].length);
-    out.push({
-      type: "mdref",
-      key: `ref${m[1]}`,
-      author: authorPart
-        .split(/,\s*|\s+and\s+/)
-        .filter(Boolean)
-        .join(" and "),
-      title: t[1].replace(/\.$/, "").trim(),
-      booktitle: rest.trim(),
-      journal: "",
-    });
-  }
-  return out;
-}
-
 function bibTextFrom(target) {
   let file = target;
   if (!existsSync(file)) die(`no such path: ${file}`);
@@ -161,16 +83,16 @@ function bibTextFrom(target) {
       }
       return undefined;
     };
-    const found =
-      pick(".bib") ||
-      pick(".tex") ||
-      names.find((f) => f === "paper.md" || f === "draft.md");
-    if (!found) die(`no .bib, .tex or paper.md in ${file}`);
+    const found = pick(".bib") || pick(".tex");
+    if (!found) die(`no .bib or .tex in ${file}`);
     file = join(file, found);
   }
   const text = readFileSync(file, "utf-8");
-  if (file.endsWith(".md")) return { text, file, markdown: true };
   if (file.endsWith(".bib")) return { text, file };
+  // Only a .bib and a .tex are read: any other file would fall through to the refs.bib beside it,
+  // and the report would name a file nobody asked about.
+  if (!file.endsWith(".tex"))
+    die(`${file} is not a .bib or a .tex — give the paper's .bib or paper.tex`);
   // A .tex may carry the bibliography inline via filecontents — that is how our papers do it.
   const m = text.match(
     /\\begin\{filecontents\*?\}(?:\[[^\]]*\])?\{[^}]*\.bib\}\r?\n([\s\S]*?)\\end\{filecontents\*?\}/,
@@ -358,13 +280,6 @@ export async function checkAuthors(
   const matched = []; // compared against the published record, and equal
   const entries = parsed.filter((e) => e.title && e.author);
 
-  for (const e of parsed.filter((x) => x.unparsed)) {
-    skipped.push({
-      key: e.key,
-      why: "no author/title — a reference to software or a dataset, DBLP does not apply",
-    });
-  }
-
   for (const e of entries) {
     if (!claimsPublished(e)) continue; // a preprint entry is allowed to carry preprint metadata
     if (truncated(e.author)) {
@@ -438,8 +353,8 @@ async function main() {
   if (!args[0])
     die("usage: bib-authors.mjs <paper-dir|file.bib|file.tex> [--json]");
 
-  const { text, file, markdown } = bibTextFrom(args[0]);
-  const parsed = markdown ? parseMarkdownRefs(text) : parseBib(text);
+  const { text, file } = bibTextFrom(args[0]);
+  const parsed = parseBib(text);
   const entries = parsed.filter((e) => e.title && e.author);
   const { findings, skipped, unchecked } = await checkAuthors(parsed);
   if (asJson) {

package/skills/verify-citations/scripts/bib-authors.test.mjs

@@ -14,12 +14,7 @@
  */
 import assert from "node:assert/strict";
 import { test } from "vitest";
-import {
-  surnames,
-  compare,
-  claimsPublished,
-  parseMarkdownRefs,
-} from "./bib-authors.mjs";
+import { surnames, compare, claimsPublished } from "./bib-authors.mjs";
 
 const t = test;
 
@@ -135,41 +130,3 @@ t("a proceedings entry IS in scope", () => {
     true,
   );
 });
-
-console.log("bib-authors — where the bibliography begins");
-
-// Both halves. "Silent on a fence" without "finds the real one" is indistinguishable from a broken parse.
-t("a real ## References is parsed", () => {
-  const md = [
-    "# Paper",
-    "text",
-    "",
-    "## References",
-    "",
-    "1. A. Author. *A title*. In NeurIPS, 2023.",
-  ].join("\n");
-  const refs = parseMarkdownRefs(md);
-  assert.equal(refs.length, 1);
-  assert.match(refs[0].title, /A title/);
-});
-
-t("a ## References inside a ``` fence does NOT open the bibliography", () => {
-  const md = [
-    "# Paper",
-    "",
-    "The section is written like this:",
-    "",
-    "```markdown",
-    "## References",
-    "",
-    "1. Z. Ghost. *Not a reference, but a markup example*. In Nowhere, 2020.",
-    "```",
-    "",
-    "The end.",
-  ].join("\n");
-  assert.deepEqual(
-    parseMarkdownRefs(md),
-    [],
-    "quoted markup is not a bibliography — the ^#+ regex counted it as a heading",
-  );
-});

package/src/build.harness.mjs

@@ -482,7 +482,7 @@ try {
   );
 
   // ── no paper.tex: a REFUSAL, not a skip ───────────────────────────────────────────────
-  const bare = paper("bare", { "paper.md": "# x" });
+  const bare = paper("bare", { "PIPELINE-STATUS.md": "# x" });
   const ntex = fakeTex();
   const nr = await buildPaper(bare, {
     cwd: root,
@@ -546,7 +546,7 @@ try {
 
   // (c) no paper.tex: refused, and a PDF left from some earlier build goes too.
   const bareStale = paper("bare-stale", {
-    "paper.md": "# x",
+    "PIPELINE-STATUS.md": "# x",
     "paper.pdf": "%PDF-stale-from-yesterday",
   });
   const bsTex = fakeTex();

package/src/build.test.ts

@@ -162,7 +162,7 @@ test("documentclass: an empty name is no class; a macro inside the name contribu
 });
 
 test("facts of a paper with no paper.tex: no class", () => {
-  assert.deepEqual(readFacts(paper({ "paper.md": "# P\n" })), {
+  assert.deepEqual(readFacts(paper({ "PIPELINE-STATUS.md": "# P\n" })), {
     main: null,
     documentclass: null,
     venue: null,

package/src/build.ts

@@ -89,12 +89,7 @@ import { messageOf, printed } from "./domain/text.ts";
 const PIPED: ["ignore", "pipe", "pipe"] = ["ignore", "pipe", "pipe"];
 
 /** A directory counts as a paper by the same markers as `structure.ts` — one shared dictionary. */
-export const PAPER_MARKERS = [
-  "PIPELINE-STATUS.md",
-  "paper.tex",
-  "paper.md",
-  CONFIG_FILE,
-];
+export const PAPER_MARKERS = ["PIPELINE-STATUS.md", "paper.tex", CONFIG_FILE];
 
 /** The source paperlint compiles, and the job name every output file carries. */
 export const MAIN = "paper.tex";