@event4u/agent-config 9.9.0 → 9.11.0

package/src/scripts/encoding_corpus_report.ts

@@ -0,0 +1,216 @@
+#!/usr/bin/env tsx
+/**
+ * Measure the encoding floor against the FROZEN corpus
+ * (road-to-runtime-encoding-hardening Phase 4).
+ *
+ * Renders the numbers from the corpus — never hand-typed — and evaluates them
+ * against the thresholds the roadmap pre-registered BEFORE the detector existed:
+ *
+ *   recall on in-scope positives            >= 0.95
+ *   recall on unambiguous classes           >= 0.99
+ *     (zero-width, bidi, variation-selector)
+ *   false-positive rate on clean negatives  <= 0.005
+ *   added runtime model spend               == 0   (structural: no model call)
+ *   added latency                           <  2 ms p95 per message
+ *
+ * A positive counts as DETECTED when the floor either strips its signal
+ * (`sanitize_text` changed the value) or reports it (`scan_encoding_findings`
+ * returned a finding). Both are "the vector did not reach the model unnoticed",
+ * which is the property under test.
+ *
+ * Usage:
+ *     encoding_corpus_report            # human summary + exit 1 if a gate fails
+ *     encoding_corpus_report --json     # machine payload
+ */
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+import { fileURLToPath, pathToFileURL } from 'node:url';
+
+import { OUT_DIR, type CorpusEntry } from './encoding_corpus.js';
+import { sanitize_text, scan_encoding_findings } from './_lib/retrieval_sanitize.js';
+
+/** Pre-registered acceptance thresholds — do not edit to make a run pass. */
+export const THRESHOLDS = {
+    recall_all: 0.95,
+    recall_unambiguous: 0.99,
+    fp_rate_max: 0.005,
+    latency_p95_ms_max: 2,
+} as const;
+
+/** Classes the roadmap singled out as unambiguous → higher recall bar. */
+export const UNAMBIGUOUS_CHANNELS: readonly string[] = [
+    'zero-width',
+    'zero-width-joiner-bom',
+    'bidi-control',
+    'variation-selector-run',
+];
+
+function _readJsonl(name: string): CorpusEntry[] {
+    return fs
+        .readFileSync(path.join(OUT_DIR, name), 'utf-8')
+        .split('\n')
+        .filter((l) => l !== '')
+        .map((l) => JSON.parse(l) as CorpusEntry);
+}
+
+/** True when the floor either stripped or reported the entry's signal. */
+export function detected(text: string): { hit: boolean; stripped: boolean; flagged: boolean } {
+    const stripped = sanitize_text(text) !== text;
+    const flagged = scan_encoding_findings(text).length > 0;
+    return { hit: stripped || flagged, stripped, flagged };
+}
+
+export interface Report {
+    readonly corpus: { positives: number; negatives: number };
+    readonly recall_all: number;
+    readonly recall_unambiguous: number;
+    readonly fp_rate: number;
+    readonly per_channel: Record<string, { n: number; detected: number; recall: number }>;
+    readonly false_positives: readonly string[];
+    readonly latency_p95_ms: number;
+    readonly model_calls: 0;
+    readonly gates: Record<string, { value: number; threshold: number; pass: boolean }>;
+    readonly verdict: 'ADOPT' | 'DETECT-AND-FLAG-ONLY' | 'FAIL';
+}
+
+function _p95(samples: readonly number[]): number {
+    const sorted = [...samples].sort((a, b) => a - b);
+    const idx = Math.min(sorted.length - 1, Math.floor(sorted.length * 0.95));
+    return sorted[idx] ?? 0;
+}
+
+export function measure(): Report {
+    const positives = _readJsonl('positives.jsonl');
+    const negatives = _readJsonl('negatives.jsonl');
+
+    const perChannel: Record<string, { n: number; detected: number; recall: number }> = {};
+    const latencies: number[] = [];
+    let hits = 0;
+    for (const e of positives) {
+        const t0 = process.hrtime.bigint();
+        const d = detected(e.text);
+        latencies.push(Number(process.hrtime.bigint() - t0) / 1e6);
+        const slot = perChannel[e.channel] ?? { n: 0, detected: 0, recall: 0 };
+        slot.n += 1;
+        if (d.hit) {
+            slot.detected += 1;
+            hits += 1;
+        }
+        slot.recall = slot.detected / slot.n;
+        perChannel[e.channel] = slot;
+    }
+
+    const falsePositives: string[] = [];
+    for (const e of negatives) {
+        const t0 = process.hrtime.bigint();
+        const d = detected(e.text);
+        latencies.push(Number(process.hrtime.bigint() - t0) / 1e6);
+        if (d.hit) falsePositives.push(`${e.id}: ${e.text.slice(0, 90)}`);
+    }
+
+    const unambiguous = positives.filter((e) => UNAMBIGUOUS_CHANNELS.includes(e.channel));
+    const unambiguousHits = unambiguous.filter((e) => detected(e.text).hit).length;
+
+    const recallAll = positives.length === 0 ? 0 : hits / positives.length;
+    const recallUnambiguous = unambiguous.length === 0 ? 0 : unambiguousHits / unambiguous.length;
+    const fpRate = negatives.length === 0 ? 0 : falsePositives.length / negatives.length;
+    const p95 = _p95(latencies);
+
+    const gates = {
+        recall_all: { value: recallAll, threshold: THRESHOLDS.recall_all, pass: recallAll >= THRESHOLDS.recall_all },
+        recall_unambiguous: {
+            value: recallUnambiguous,
+            threshold: THRESHOLDS.recall_unambiguous,
+            pass: recallUnambiguous >= THRESHOLDS.recall_unambiguous,
+        },
+        fp_rate: { value: fpRate, threshold: THRESHOLDS.fp_rate_max, pass: fpRate <= THRESHOLDS.fp_rate_max },
+        latency_p95_ms: {
+            value: p95,
+            threshold: THRESHOLDS.latency_p95_ms_max,
+            pass: p95 < THRESHOLDS.latency_p95_ms_max,
+        },
+    };
+
+    // The roadmap's pre-registered branch: recall clears but FP is over budget →
+    // ship detect-and-flag only, never automatic quarantine, and publish the FP
+    // number. Nothing here quarantines anyway, so this records which branch the
+    // measurement selected rather than switching behaviour after the fact.
+    const recallOk = gates.recall_all.pass && gates.recall_unambiguous.pass;
+    const verdict: Report['verdict'] = !recallOk
+        ? 'FAIL'
+        : gates.fp_rate.pass && gates.latency_p95_ms.pass
+          ? 'ADOPT'
+          : 'DETECT-AND-FLAG-ONLY';
+
+    return {
+        corpus: { positives: positives.length, negatives: negatives.length },
+        recall_all: recallAll,
+        recall_unambiguous: recallUnambiguous,
+        fp_rate: fpRate,
+        per_channel: perChannel,
+        false_positives: falsePositives,
+        latency_p95_ms: p95,
+        model_calls: 0,
+        gates,
+        verdict,
+    };
+}
+
+function _render(r: Report): string {
+    const pct = (n: number): string => (n * 100).toFixed(2) + '%';
+    const lines: string[] = [];
+    lines.push(`corpus: ${r.corpus.positives} positives · ${r.corpus.negatives} negatives`);
+    lines.push('');
+    lines.push('per-channel recall');
+    for (const [ch, s] of Object.entries(r.per_channel).sort()) {
+        const mark = s.recall === 1 ? '  ' : '!!';
+        lines.push(`  ${mark} ${ch.padEnd(24)} ${s.detected}/${s.n}  ${pct(s.recall)}`);
+    }
+    lines.push('');
+    for (const [name, g] of Object.entries(r.gates)) {
+        const cmp = name === 'fp_rate' || name === 'latency_p95_ms' ? '<=' : '>=';
+        const val = name === 'latency_p95_ms' ? `${g.value.toFixed(3)} ms` : pct(g.value);
+        const thr = name === 'latency_p95_ms' ? `${g.threshold} ms` : pct(g.threshold);
+        lines.push(`  ${g.pass ? 'PASS' : 'FAIL'}  ${name.padEnd(20)} ${val} ${cmp} ${thr}`);
+    }
+    lines.push(`  PASS  ${'model_calls'.padEnd(20)} 0 (structural: no model call on this path)`);
+    if (r.false_positives.length > 0) {
+        lines.push('');
+        lines.push(`false positives (${r.false_positives.length}):`);
+        for (const fp of r.false_positives.slice(0, 12)) lines.push(`  - ${fp}`);
+        if (r.false_positives.length > 12) {
+            lines.push(`  … ${r.false_positives.length - 12} more`);
+        }
+    }
+    lines.push('');
+    lines.push(`verdict: ${r.verdict}`);
+    return lines.join('\n') + '\n';
+}
+
+export function main(argv: readonly string[] | null = null): number {
+    const args = argv ?? process.argv.slice(2);
+    const r = measure();
+    if (args.includes('--json')) {
+        process.stdout.write(JSON.stringify(r, null, 2) + '\n');
+    } else {
+        process.stdout.write(_render(r));
+    }
+    return r.verdict === 'FAIL' ? 1 : 0;
+}
+
+declare const __AGENT_CONFIG_BUNDLE__: boolean | undefined;
+function _isCliEntry(): boolean {
+    if (typeof __AGENT_CONFIG_BUNDLE__ !== 'undefined' && __AGENT_CONFIG_BUNDLE__) return false;
+    if (process.argv[1] === undefined) return false;
+    const argvUrl = pathToFileURL(path.resolve(process.argv[1])).href;
+    if (import.meta.url === argvUrl) return true;
+    try {
+        return fs.realpathSync(fileURLToPath(import.meta.url)) === fs.realpathSync(path.resolve(process.argv[1]));
+    } catch {
+        return false;
+    }
+}
+
+if (_isCliEntry()) {
+    process.exitCode = main();
+}

package/src/scripts/evaluator_umbrella.sh

@@ -65,11 +65,38 @@ UNPACKED_BYTES="$(npm pack --dry-run --json --ignore-scripts --silent 2>/dev/nul
 step "headless consumer install from the tarball"
 WORK="$(mktemp -d)"
 trap 'rm -rf "$WORK"' EXIT
+COLD_START_T0="$(node -e 'console.log(Date.now())')"
 ( cd "$WORK" && npm init -y >/dev/null 2>&1 \
   && npm install --omit=dev --no-audit --no-fund --silent "$ROOT/$TARBALL" )
 PKG_DIR="$WORK/node_modules/@event4u/agent-config"
 [[ -d "$PKG_DIR" ]] || { echo "❌ install did not produce the package dir" >&2; exit 1; }
 
+# ------------------------------------------- 3b. cold start: install → doctor
+# The bare-machine wall-clock a first-time user lives through, measured end to
+# end from `npm install` to `doctor` having run. This is EVIDENCE, never a
+# promise, and deliberately NOT a budget entry: it is dominated by network and
+# registry latency, which are not ours to guarantee. A gate on it would flap on
+# runner weather and teach the reader to ignore the line
+# (road-to-zero-ceremony-install Phase 6).
+step "cold start: install → doctor (evidence, not a gate)"
+COLD_START_MS="$(node -e '
+const {execFileSync}=require("node:child_process");
+const [bin, t0] = process.argv.slice(1);
+try { execFileSync(bin, ["doctor"], {stdio:"ignore"}); }
+catch { /* doctor may report findings; the run still bounds the wall-clock */ }
+console.log(Date.now() - Number(t0));' "$WORK/node_modules/.bin/agent-config" "$COLD_START_T0")"
+COLD_START_CONDITIONS="$(uname -s)/$(uname -m) · node $(node --version) · npm $(npm --version) · npm install --omit=dev from a LOCAL tarball (no registry round-trip for the package itself; a real first-touch install adds registry latency)"
+printf 'cold_start_install_to_doctor_ms=%s\n' "$COLD_START_MS"
+printf 'cold_start_conditions=%s\n' "$COLD_START_CONDITIONS"
+if [[ -n "${GITHUB_STEP_SUMMARY:-}" ]]; then
+  {
+    printf '### Cold start — install → doctor\n\n'
+    printf '**%s ms**\n\n' "$COLD_START_MS"
+    printf 'Conditions: %s\n\n' "$COLD_START_CONDITIONS"
+    printf 'Evidence, not a promise — no shipped claim states this as a guarantee.\n'
+  } >> "$GITHUB_STEP_SUMMARY"
+fi
+
 # --------------------------------------------------------- 4. staleness lints
 step "staleness lints (pre-migration refs)"
 ./scripts-run src/scripts/lint_pre_migration_refs
@@ -128,10 +155,16 @@ const m={
   mcp_boot_to_initialize_ms: Number(process.argv[6]),
   cli_help_command_count: Number(process.argv[7]),
   mcp_public_tool_count: Number(process.argv[8]),
+  // Evidence-only: carried in the artifact with its conditions, never gated.
+  // check_evaluator_budgets iterates the BUDGET entries, so an extra
+  // measurement is recorded and ignored by the gate — by design.
+  cold_start_install_to_doctor_ms: Number(process.argv[9]),
+  cold_start_conditions: process.argv[10],
 };
 fs.writeFileSync(process.argv[1], JSON.stringify(m, null, 2));
 console.log(JSON.stringify(m, null, 1));' \
-  "$MEASURE_TMP" "$UNPACKED_BYTES" "$NM_MB" "$DEP_COUNT" "$CLI_MS" "$MCP_MS" "$CLI_COUNT" "$MCP_TOOLS"
+  "$MEASURE_TMP" "$UNPACKED_BYTES" "$NM_MB" "$DEP_COUNT" "$CLI_MS" "$MCP_MS" "$CLI_COUNT" "$MCP_TOOLS" \
+  "$COLD_START_MS" "$COLD_START_CONDITIONS"
 
 step "budget gate (absolute + >regression_pct creep)"
 ./scripts-run src/scripts/check_evaluator_budgets --measurements "$MEASURE_TMP"

package/src/scripts/generate_index.ts

@@ -25,6 +25,7 @@ import {
     strip_source_prefix,
 } from './_lib/agent_src.js';
 import { computeStatus as domainSoundnessStatus } from './domain_soundness_status.js';
+import { skip_compile_disabled_rule } from './condense.js';
 
 const _HERE = fileURLToPath(import.meta.url);
 
@@ -243,6 +244,12 @@ export function _collect_skills(): Entry[] {
     return _sortedKeys(seen).map((k) => seen.get(k) as Entry);
 }
 
+// docs/catalog.md is a SHIPPED file, so a relative link from it must resolve inside
+// package.json#files. docs/proof.md is not shipped and cannot be: it is 58KB of
+// generated output linking eight paths under internal/bench/, tests/ and src/scripts/
+// that are themselves unshipped, so shipping it would hand consumers a page of dead
+// links. The published site URL is the honest target for a consumer-facing pointer.
+
 export function _collect_rules(): Entry[] {
     const seen = new Map<string, Entry>();
     for (const srcRoot of artefact_roots()) {
@@ -255,6 +262,14 @@ export function _collect_rules(): Entry[] {
             if (seen.has(stem)) {
                 continue;
             }
+            // A compile-disabled rule has no dist/agent-src/ counterpart, and the
+            // table links the SHIPPED path — so listing it emits a dead link. The
+            // toggle is a separate axis from the source file's existence; this is its
+            // fifth consumer after the router compiler, the dist writer, check_sync,
+            // and the per-tool symlink generator. Same predicate in all five.
+            if (skip_compile_disabled_rule(`rules/${stem}.md`)) {
+                continue;
+            }
             const fm = _parse_frontmatter(fs.readFileSync(ruleMd, 'utf-8'));
             seen.set(stem, {
                 kind: 'rule',
@@ -420,7 +435,8 @@ export function _render_catalog(
         '> not yet. Coverage is reported per tier (`./scripts-run',
         '> src/scripts/skill_eval_coverage`) and CI-ratcheted so it can only rise.',
         '> Listing here never implies a skill is behaviourally tested — the',
-        '> current coverage and its gap are published on [the proof page](proof.md).',
+        '> current coverage and its gap are published on [the proof page]' +
+            '(https://event4u-app.github.io/agent-config/proof/).',
         '',
         '> **Non-coding domain correctness is scoped, not proven.** The',
         '> `finance` / `founder` / `ops` / `content` skills are forged on TS/PHP',
@@ -429,7 +445,8 @@ export function _render_catalog(
         '> cited formulas, or rubric targets matching a named external practice);',
         '> the rest are labeled `unvalidated` until they pass one',
         '> (`./scripts-run src/scripts/domain_soundness_status`). A disclaimer',
-        '> floor bounds liability, not correctness — see [the proof page](proof.md).',
+        '> floor bounds liability, not correctness — see [the proof page]' +
+            '(https://event4u-app.github.io/agent-config/proof/).',
         '',
         _render_table(skills, ['kind', 'name', 'source', 'description'], '../', _to_shipped_path),
         '',

package/src/scripts/install-hooks.sh

@@ -30,8 +30,8 @@ cat > "$HOOKS_DIR/pre-push" << 'EOF'
 # the old hook only verified dist sync + the COMMAND count — it missed
 # guideline/skill/rule count drift and the generate-tools / router / corpus
 # outputs (e.g. a new guideline bumping the count in README + docs/architecture).
-# `task consistency` is the EXACT CI mirror (sync-check + sync-check-hashes +
-# sync + generate-tools + compile-router + compile-corpus + `git diff --quiet`),
+# `task consistency` is the EXACT CI mirror (sync-check + sync + generate-tools
+# + compile-router + compile-corpus + `git diff --quiet`),
 # so any derived-output drift blocks the push here. Runtime ~15-40s (it
 # regenerates the tool trees) — cheaper than a red CI run and a fixup re-push.
 

package/src/scripts/lint_confusables.ts

@@ -34,86 +34,24 @@ import * as fs from 'node:fs';
 import { fileURLToPath, pathToFileURL } from 'node:url';
 
 import * as sl from './_lib/security_lint.js';
+import { TOKEN_RE, classifyToken } from './_lib/confusables.js';
 
 export const CHECK = 'mixed-script-confusable';
 
-const MIN_LETTERS = 3;
-
-// Unicode-script classifiers via property escapes (requires the `u` flag).
-const _LATIN = /\p{Script=Latin}/u;
-const _CYRILLIC = /\p{Script=Cyrillic}/u;
-const _GREEK = /\p{Script=Greek}/u;
-const _LETTER = /\p{L}/u;
-// A "token" is a maximal run of letters / combining marks / decimal digits.
-const _TOKEN = /[\p{L}\p{M}\p{Nd}]+/gu;
-
-// Cyrillic + Greek codepoints that have a basic-Latin lookalike (Unicode TR39
-// confusables, common subset). A foreign letter only counts toward the
-// confusable signal when it is in this set — Greek math operators with no Latin
-// twin (Δ Σ Π Ω Φ Θ Λ Ξ Ψ Γ δ π ω …) are deliberately excluded so legit
-// notation does not false-positive.
-const _CONFUSABLE_FOREIGN: ReadonlySet<number> = new Set([
-    // Cyrillic lowercase ↔ Latin
-    0x0430 /*а→a*/, 0x0435 /*е→e*/, 0x043e /*о→o*/, 0x0440 /*р→p*/, 0x0441 /*с→c*/,
-    0x0443 /*у→y*/, 0x0445 /*х→x*/, 0x0455 /*ѕ→s*/, 0x0456 /*і→i*/, 0x0458 /*ј→j*/,
-    0x04bb /*һ→h*/, 0x0501 /*ԁ→d*/, 0x051b /*ԛ→q*/, 0x0577 /*no*/,
-    // Cyrillic uppercase ↔ Latin
-    0x0410 /*А→A*/, 0x0412 /*В→B*/, 0x0415 /*Е→E*/, 0x041a /*К→K*/, 0x041c /*М→M*/,
-    0x041d /*Н→H*/, 0x041e /*О→O*/, 0x0420 /*Р→P*/, 0x0421 /*С→C*/, 0x0422 /*Т→T*/,
-    0x0423 /*У→Y*/, 0x0425 /*Х→X*/, 0x0405 /*Ѕ→S*/, 0x0406 /*І→I*/, 0x0408 /*Ј→J*/,
-    // Greek lowercase ↔ Latin
-    0x03b1 /*α→a*/, 0x03b5 /*ε→e*/, 0x03bf /*ο→o*/, 0x03bd /*ν→v*/, 0x03c1 /*ρ→p*/,
-    0x03c5 /*υ→u*/, 0x03c7 /*χ→x*/, 0x03ba /*κ→k*/,
-    // Greek uppercase ↔ Latin (only those with a real Latin twin)
-    0x0391 /*Α→A*/, 0x0392 /*Β→B*/, 0x0395 /*Ε→E*/, 0x0396 /*Ζ→Z*/, 0x0397 /*Η→H*/,
-    0x0399 /*Ι→I*/, 0x039a /*Κ→K*/, 0x039c /*Μ→M*/, 0x039d /*Ν→N*/, 0x039f /*Ο→O*/,
-    0x03a1 /*Ρ→P*/, 0x03a4 /*Τ→T*/, 0x03a5 /*Υ→Y*/, 0x03a7 /*Χ→X*/, 0x0392 /*Β→B*/,
-]);
-
-type ScriptName = 'latin' | 'cyrillic' | 'greek' | 'other';
-
-function _script_of(ch: string): ScriptName {
-    if (_LATIN.test(ch)) return 'latin';
-    if (_CYRILLIC.test(ch)) return 'cyrillic';
-    if (_GREEK.test(ch)) return 'greek';
-    return 'other';
-}
+// The confusable table + the mixed-script decision rule live in ONE place so
+// this authoring-time linter and the runtime encoding scanner cannot drift apart
+// (road-to-runtime-encoding-hardening Phase 3). `_classify_token` stays exported
+// under its original name — it is this module's published surface and its tests
+// address it here.
+const _TOKEN = TOKEN_RE;
 
 /**
  * Classify a single token. Returns the offending mixed-script descriptor when
  * the token matches the homoglyph signature, else null.
+ *
+ * Thin re-export of the shared implementation — see `_lib/confusables.ts`.
  */
-export function _classify_token(token: string): string | null {
-    let latin = 0;
-    let cyrillic = 0;
-    let greek = 0;
-    let letters = 0;
-    for (const ch of token) {
-        if (!_LETTER.test(ch)) continue; // skip combining marks / digits for script vote
-        letters += 1;
-        const script = _script_of(ch);
-        if (script === 'latin') {
-            latin += 1;
-        } else if (script === 'cyrillic' || script === 'greek') {
-            // Only confusable foreign letters (TR39) count — math operators with
-            // no Latin twin are ignored, so legit notation never trips.
-            if (_CONFUSABLE_FOREIGN.has(ch.codePointAt(0) as number)) {
-                if (script === 'cyrillic') cyrillic += 1;
-                else greek += 1;
-            }
-        }
-    }
-    if (letters < MIN_LETTERS) return null;
-    const foreign = cyrillic + greek;
-    if (latin === 0 || foreign === 0) return null; // single-script (or no Latin) → legit
-    if (latin <= foreign) return null; // majority must be Latin (foreign word ≠ attack)
-    const which = cyrillic > 0 && greek > 0
-        ? 'cyrillic+greek'
-        : cyrillic > 0
-            ? 'cyrillic'
-            : 'greek';
-    return `latin+${which} (${latin} latin / ${foreign} foreign-confusable)`;
-}
+export const _classify_token = classifyToken;
 
 export function _scan(sf: sl.ScannedFile): sl.Finding[] {
     if (sf.pragma_allows(CHECK)) {

package/src/scripts/lint_eval_fixture_citations.ts

@@ -0,0 +1,113 @@
+#!/usr/bin/env tsx
+/**
+ * Every design-artifact eval fixture id must be cited by some surface.
+ *
+ * The fixtures file used to assert that the lifecycle contract's branch table
+ * "references these ids". It cites nine of them; the rest are gated elsewhere
+ * (the design-fidelity mechanics guideline, the lane-matrix test) — and a
+ * handful were cited by nothing at all. An uncited fixture is an eval nobody
+ * runs and no contract depends on, which is exactly the shape of coverage that
+ * looks real in a count and is absent in practice.
+ *
+ * This checks citation, not correctness: an id must appear outside its own
+ * definition. Which surface cites it is the author's call.
+ *
+ * Exit codes: 0 every id cited, 1 one or more uncited.
+ */
+
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const QUIET = process.argv.includes('--quiet');
+
+const _HERE = fileURLToPath(import.meta.url);
+const REPO = path.resolve(path.dirname(_HERE), '..', '..');
+const FIXTURES = path.join(REPO, 'tests', 'design-artifacts', 'eval-fixtures.md');
+
+/** Surfaces allowed to satisfy a citation. */
+const SCAN_ROOTS: readonly string[] = [
+    path.join(REPO, 'docs', 'contracts'),
+    path.join(REPO, 'docs', 'guidelines'),
+    path.join(REPO, 'src', 'skills'),
+    path.join(REPO, 'src', 'rules'),
+    path.join(REPO, 'tests', 'scripts'),
+];
+
+const SCAN_EXTS: ReadonlySet<string> = new Set(['.md', '.ts']);
+
+function walk(dir: string, out: string[]): void {
+    let entries: fs.Dirent[];
+    try {
+        entries = fs.readdirSync(dir, { withFileTypes: true });
+    } catch {
+        return;
+    }
+    for (const entry of entries) {
+        const p = path.join(dir, entry.name);
+        if (entry.isDirectory()) {
+            walk(p, out);
+        } else if (SCAN_EXTS.has(path.extname(entry.name))) {
+            out.push(p);
+        }
+    }
+}
+
+function main(): number {
+    let fixtureText: string;
+    try {
+        fixtureText = fs.readFileSync(FIXTURES, { encoding: 'utf-8' });
+    } catch {
+        process.stderr.write(`lint-eval-fixture-citations: cannot read ${FIXTURES}\n`);
+        return 1;
+    }
+
+    // Ids are declared as `### daf-...` headings or as leading table cells.
+    const ids = new Set<string>();
+    for (const m of fixtureText.matchAll(/^###\s+(daf-[a-z0-9-]+)\s*$/gm)) {
+        ids.add(m[1] as string);
+    }
+    for (const m of fixtureText.matchAll(/^\|\s*`(daf-[a-z0-9-]+)`\s*\|/gm)) {
+        ids.add(m[1] as string);
+    }
+    if (ids.size === 0) {
+        process.stderr.write('lint-eval-fixture-citations: no fixture ids found — parser drift.\n');
+        return 1;
+    }
+
+    const files: string[] = [];
+    for (const root of SCAN_ROOTS) {
+        walk(root, files);
+    }
+    const haystack = files
+        .map((f) => {
+            try {
+                return fs.readFileSync(f, { encoding: 'utf-8' });
+            } catch {
+                return '';
+            }
+        })
+        .join('\n');
+
+    const uncited = [...ids].filter((id) => !haystack.includes(id)).sort();
+    if (uncited.length > 0) {
+        for (const id of uncited) {
+            process.stdout.write(
+                `lint-eval-fixture-citations: \`${id}\` is cited by no contract, ` +
+                    'guideline, skill, rule, or test — an eval nobody runs.\n',
+            );
+        }
+        process.stderr.write(
+            `lint-eval-fixture-citations: ${uncited.length} of ${ids.size} fixture id(s) uncited.\n`,
+        );
+        return 1;
+    }
+    if (!QUIET) {
+        process.stdout.write(
+            `lint-eval-fixture-citations: ${ids.size} fixture id(s), all cited.\n`,
+        );
+    }
+    return 0;
+}
+
+process.exit(main());