@dogfood-lab/verify 1.7.0 → 1.8.0

package/README.md

@@ -62,6 +62,28 @@ import { validateSchema } from '@dogfood-lab/verify/validators/schema.js';
 import { validateProvenance } from '@dogfood-lab/verify/validators/provenance.js';
 ```
 
+## CLI (`dogfood-verify`)
+
+The package ships a `dogfood-verify` bin with two verbs.
+
+**Verify** a submission (local dry-run / explain — never writes):
+
+```bash
+dogfood-verify --file submission.json --explain   # human verdict, reasons classified by who-fixes-it
+dogfood-verify --file submission.json --json       # machine-readable
+# exit: 0 accepted · 1 rejected · 2 operator error
+```
+
+**Lint** a policy file (VERIFY-F3, author-time — no submission needed):
+
+```bash
+dogfood-verify lint policies/repos/<org>/<repo>.yaml        # or global-policy.yaml
+dogfood-verify lint policies/global-policy.yaml --json      # for CI
+# exit: 0 clean or warnings-only · 1 errors · 2 operator error
+```
+
+`lint` runs the structural schema gate **plus** the data-independent predicate checks (`unknown_field`, `max_depth`, `node_budget`) over every `when` predicate, and emits an **advisory** warning on the `[]` footgun (a negative operator over a `[]` path fails open) with the fail-closed `not(any(...))` rewrite as a suggestion — never auto-applied, never a hard error. It is the `opa check` analogue. It **cannot** statically catch a `type_mismatch` or a fan-out overrun (both are data-dependent) and says so. Full contract + coverage boundary: [`docs/policy-lint.md`](https://github.com/dogfood-lab/testing-os/blob/main/docs/policy-lint.md).
+
 ## Submission envelope
 
 The full envelope shape is defined by `@dogfood-lab/schemas` (`dogfood-record-submission.schema.json`). Minimum required fields:

package/cli-lint.js

@@ -0,0 +1,209 @@
+/**
+ * cli-lint.js (VERIFY-F3) — the `dogfood-verify lint <policy-file>` subcommand.
+ *
+ * A SEPARATE parse/render path from the verify CLI (cli.js): it takes a policy YAML
+ * (not a submission JSON) and reports static lint findings, so it does not share the
+ * verify arg parser or output. cli.js's `main` dispatcher routes the `lint` verb here
+ * and leaves the verify `run` path untouched. The exit contract mirrors that path:
+ *
+ *   0 — clean, or warnings-only (footgun advisories never block).
+ *   1 — one or more errors (schema-invalid, a static predicate fault, or unparseable YAML).
+ *   2 — operator error (file missing/unreadable, or a malformed invocation).
+ *
+ * YAML that fails to parse is exit 1 (a lint FINDING about the policy the author must fix —
+ * surfacing "line 4: bad indentation" is the lint's job), not exit 2. A file that does not
+ * exist is exit 2 (the author pointed at the wrong path). See docs/policy-lint.md.
+ */
+
+import { readFileSync } from 'node:fs';
+import { resolve } from 'node:path';
+import yaml from 'js-yaml';
+
+import { lintPolicy, COVERAGE_NOTE } from './validators/lint-policy.js';
+
+/** Operator-error sentinel → exit 2 (distinct from a lint finding, which is exit 1). */
+class LintOperatorError extends Error {
+  constructor(message, hint) {
+    super(message);
+    this.name = 'LintOperatorError';
+    this.hint = hint;
+  }
+}
+
+const LINT_USAGE = `dogfood-verify lint — author-time static check for a policy file
+
+USAGE:
+  dogfood-verify lint <policy-file> [--json]
+
+WHAT IT CHECKS (no submission needed):
+  - structural validity against policy.schema.json
+  - every predicate's known leading field, combinator depth, and node budget
+  - an ADVISORY warning on the [] footgun (a negative op over a [] path fails open)
+
+It CANNOT statically catch a type_mismatch or a fanout_budget overrun — those depend
+on submission data. Run \`dogfood-verify --file <submission> --explain\` for that.
+
+OPTIONS:
+  --json        Machine-readable result for CI.
+  -h, --help    Show this help.
+
+EXIT CODES:
+  0  clean or warnings-only     1  errors found     2  operator error (bad flags / IO)`;
+
+/**
+ * Parse the lint argv (everything AFTER the `lint` verb). Accepts exactly one positional
+ * policy-file path plus optional `--json` / `--help`. Throws LintOperatorError (→ exit 2)
+ * on any malformed invocation.
+ *
+ * @param {string[]} argv
+ * @returns {{ help: boolean, file: string|null, json: boolean }}
+ */
+export function parseLintArgs(argv) {
+  let file = null;
+  let json = false;
+  let help = false;
+
+  for (const arg of argv) {
+    if (arg === '-h' || arg === '--help') { help = true; continue; }
+    if (arg === '--json') { json = true; continue; }
+    if (arg.startsWith('-')) {
+      throw new LintOperatorError(`unknown argument: ${arg}`, 'run `dogfood-verify lint --help` for usage');
+    }
+    if (file !== null) {
+      throw new LintOperatorError('more than one policy file given', 'lint one file at a time');
+    }
+    file = arg;
+  }
+
+  if (help) return { help: true, file: null, json: false };
+  if (file === null) {
+    throw new LintOperatorError('no policy file provided', 'dogfood-verify lint <policy-file>');
+  }
+  return { help: false, file, json };
+}
+
+/**
+ * Classify a policy file by its path so the report can name the origin (which decides the
+ * runtime fault class: global → operational, repo → submission-bad). Mirrors the
+ * `policies/global-policy.yaml` vs `policies/repos/<org>/<repo>.yaml` layout.
+ */
+export function originForPath(p) {
+  const norm = String(p).replace(/\\/g, '/');
+  if (/\/policies\/repos\//.test(norm)) return 'repo';
+  if (/(^|\/)global-policy\.yaml$/.test(norm)) return 'global';
+  return 'unknown';
+}
+
+/** Render the human (default) view of a lint result — verdict-first, ERROR before WARNING. */
+export function renderLintText(result, file) {
+  const lines = [];
+  const verdict = !result.ok ? 'ERRORS' : (result.warnings.length ? 'CLEAN (advisory warnings)' : 'CLEAN');
+  lines.push(`VERDICT: ${verdict}`);
+  lines.push('');
+  lines.push(`  file:   ${file}`);
+  lines.push(`  origin: ${result.origin}`);
+
+  if (result.errors.length) {
+    lines.push('');
+    lines.push(`ERRORS (${result.errors.length}):`);
+    for (const e of result.errors) {
+      const field = e.field ? ` — field "${e.field}"` : '';
+      lines.push(`  - [${e.label} ${e.code}] ${e.location}${field}`);
+      lines.push(`      ${e.message}`);
+    }
+  }
+
+  if (result.warnings.length) {
+    lines.push('');
+    lines.push(`WARNINGS (${result.warnings.length}) — advisory; the author confirms intent, nothing is auto-applied:`);
+    for (const w of result.warnings) {
+      lines.push(`  - [${w.label} ${w.code}] ${w.location} — field "${w.field}"`);
+      lines.push(`      ${w.message}`);
+      lines.push(`      ${w.suggestion}`);
+    }
+  }
+
+  if (result.ok && !result.warnings.length) {
+    lines.push('');
+    lines.push('No findings. The policy passes static lint.');
+  }
+
+  lines.push('');
+  lines.push(`note: ${result.coverageNote}`);
+  return lines.join('\n');
+}
+
+/** Build the machine-readable (--json) result. */
+export function buildLintJson(result, file) {
+  return {
+    file,
+    origin: result.origin,
+    ok: result.ok,
+    errors: result.errors,
+    warnings: result.warnings,
+    coverageNote: result.coverageNote,
+  };
+}
+
+/**
+ * Run the lint subcommand. Returns the exit code (does not call process.exit) so it is
+ * unit-testable with an injected stdout/stderr sink.
+ *
+ * @param {string[]} argv - args AFTER the `lint` verb
+ * @param {{ stdout?: (s: string) => void, stderr?: (s: string) => void }} [io]
+ * @returns {Promise<number>}
+ */
+export async function runLint(argv, io = {}) {
+  const out = io.stdout ?? ((s) => process.stdout.write(s + '\n'));
+  const err = io.stderr ?? ((s) => process.stderr.write(s + '\n'));
+
+  let opts;
+  try {
+    opts = parseLintArgs(argv);
+  } catch (e) {
+    err(`ERROR: ${e.message}`);
+    if (e.hint) err(`  hint: ${e.hint}`);
+    return 2;
+  }
+
+  if (opts.help) {
+    out(LINT_USAGE);
+    return 0;
+  }
+
+  const path = resolve(opts.file);
+  let raw;
+  try {
+    raw = readFileSync(path, 'utf-8');
+  } catch (e) {
+    err(`ERROR: could not read policy file: ${path} — ${e.message}`);
+    err('  hint: check the path exists and is readable');
+    return 2;
+  }
+
+  let doc;
+  try {
+    doc = yaml.load(raw);
+  } catch (e) {
+    // A YAML parse failure is a lint finding about the policy (exit 1), not an operator error.
+    const where = e && e.mark ? ` at line ${e.mark.line + 1}, column ${e.mark.column + 1}` : '';
+    const result = {
+      ok: false,
+      origin: originForPath(path),
+      errors: [{
+        label: 'policy-schema:',
+        code: 'yaml_parse',
+        location: '/',
+        message: `policy YAML failed to parse${where} — ${e.message}`,
+      }],
+      warnings: [],
+      coverageNote: COVERAGE_NOTE,
+    };
+    out(opts.json ? JSON.stringify(buildLintJson(result, path)) : renderLintText(result, path));
+    return 1;
+  }
+
+  const result = lintPolicy(doc, { origin: originForPath(path) });
+  out(opts.json ? JSON.stringify(buildLintJson(result, path)) : renderLintText(result, path));
+  return result.ok ? 0 : 1;
+}

package/cli.js

@@ -39,6 +39,7 @@ import yaml from 'js-yaml';
 
 import { verify, parseRejectionReason } from './index.js';
 import { stubProvenance, provenanceForProvider } from './validators/provenance.js';
+import { runLint } from './cli-lint.js';
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
 
@@ -440,8 +441,27 @@ export async function run(argv, io = {}) {
   return record.verification?.status === 'accepted' ? 0 : 1;
 }
 
+/**
+ * Top-level dispatcher. The bin has two verbs:
+ *   - `dogfood-verify lint <policy-file>` → the author-time policy lint (VERIFY-F3, cli-lint.js).
+ *   - `dogfood-verify <flags>`            → the verify dry-run/explain (the original `run` path).
+ *
+ * Dispatching on `argv[0] === 'lint'` is a purely additive change: that token previously hit the
+ * verify parser's default arm and threw `unknown argument: lint`. The verify path is unchanged.
+ *
+ * @param {string[]} argv - process.argv.slice(2)
+ * @param {object} [io] - injected stdout/stderr/repoRoot for testing
+ * @returns {Promise<number>} exit code
+ */
+export async function main(argv, io = {}) {
+  if (argv[0] === 'lint') {
+    return runLint(argv.slice(1), io);
+  }
+  return run(argv, io);
+}
+
 // --- CLI entrypoint ---
 const isMain = process.argv[1] && resolve(process.argv[1]) === resolve(__dirname, 'cli.js');
 if (isMain) {
-  run(process.argv.slice(2)).then((code) => process.exit(code));
+  main(process.argv.slice(2)).then((code) => process.exit(code));
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dogfood-lab/verify",
-  "version": "1.7.0",
+  "version": "1.8.0",
   "type": "module",
   "description": "Central verifier for testing-os. Validates submissions against schema and policy, produces persisted records.",
   "main": "index.js",
@@ -10,6 +10,7 @@
   "exports": {
     ".": "./index.js",
     "./cli.js": "./cli.js",
+    "./cli-lint.js": "./cli-lint.js",
     "./parse-rejection.js": "./parse-rejection.js",
     "./validators/*": "./validators/*",
     "./validators/*.js": "./validators/*.js"
@@ -21,6 +22,7 @@
   "files": [
     "index.js",
     "cli.js",
+    "cli-lint.js",
     "parse-rejection.js",
     "validators/",
     "README.md",

package/validators/lint-policy.js

@@ -0,0 +1,120 @@
+/**
+ * Policy lint (VERIFY-F3)
+ *
+ * `lintPolicy(policyDoc, { origin })` is the author-time check behind the
+ * `dogfood-verify lint` verb: it validates a whole policy file WITHOUT a submission,
+ * batch-reporting every fault. It is the `opa check` analogue named (as a deferred
+ * companion) in docs/policy-dsl.md and specified in docs/policy-lint.md.
+ *
+ * Three passes, mirroring the runtime's own layering:
+ *   1. Structural gate — `validatePayload('policy', …)` against policy.schema.json (the same
+ *      gate loadGlobalPolicy / loadRepoPolicy run). Catches unknown op, malformed/banned field,
+ *      mixed node, value-arity, custom_rules-under-defaults, additionalProperties.
+ *   2. Static predicate walk — the data-independent semantic faults the schema cannot express:
+ *      unknown leading field, combinator over-depth, node budget (via predicate.js#lintPredicate).
+ *   3. `[]`-footgun advisory — a deterministic warning (never an error) on a negative op over a
+ *      `[]` path (via predicate.js#findEmptyArrayFootguns).
+ *
+ * Coverage boundary (the VERIFY-F2 over-claim lesson): `type_mismatch` and `fanout_budget` are
+ * DATA-DEPENDENT and cannot be caught statically. `coverageNote` says so — a clean lint is "no
+ * static fault and no footgun," NOT "this policy can never produce a policy-config: rejection."
+ */
+
+import { validatePayload } from '@dogfood-lab/schemas';
+import { lintPredicate, findEmptyArrayFootguns } from './predicate.js';
+
+/** Stated in every lint result so a clean verdict is never read as full coverage. */
+export const COVERAGE_NOTE =
+  'Static lint only — it cannot catch a `type_mismatch` (a numeric op over a non-number field) ' +
+  'or a `fanout_budget` overrun; both depend on submission data. Run ' +
+  '`dogfood-verify --file <submission> --explain` to exercise the data-dependent path.';
+
+/**
+ * Lint a parsed policy document.
+ *
+ * @param {unknown} policyDoc - The parsed policy YAML/JSON (any value; a non-object is reported
+ *   by the schema gate).
+ * @param {{ origin?: 'global'|'repo'|'unknown' }} [opts] - The policy's origin, used only for
+ *   reporting (the caller derives it from the file path).
+ * @returns {{
+ *   ok: boolean, origin: string, coverageNote: string,
+ *   errors: { label: string, code: string, location: string, field?: string, message: string }[],
+ *   warnings: { label: string, code: string, location: string, field: string, suggestion: string, message: string }[]
+ * }} `ok` is true iff there are no errors; warnings (footguns) never affect `ok`.
+ */
+export function lintPolicy(policyDoc, { origin = 'unknown' } = {}) {
+  const errors = [];
+  const warnings = [];
+
+  // 1. Structural schema gate.
+  const schema = validatePayload('policy', policyDoc);
+  for (const e of schema.errors) {
+    errors.push({
+      label: 'policy-schema:',
+      code: e.keyword || 'schema',
+      location: e.path || '/',
+      message: e.message || 'schema violation',
+    });
+  }
+
+  // 2 + 3. Static predicate walk + footgun advisory over every `when` location. Defensive: a
+  // schema-invalid doc may have malformed predicates; the walkers tolerate non-objects.
+  for (const { when, scope, location } of collectPredicates(policyDoc)) {
+    for (const f of lintPredicate(when, scope)) {
+      errors.push({ label: 'policy-config:', code: f.code, location, field: f.field, message: f.message });
+    }
+    for (const g of findEmptyArrayFootguns(when)) {
+      warnings.push({
+        label: 'policy-footgun:',
+        code: 'footgun-empty-array',
+        location,
+        field: g.field,
+        suggestion: g.suggestion,
+        message:
+          `negative operator "${g.op}" over the array path "${g.field}" fails OPEN on an empty or ` +
+          `absent array — the rule silently does not fire when there are no elements`,
+      });
+    }
+  }
+
+  return { ok: errors.length === 0, origin, errors, warnings, coverageNote: COVERAGE_NOTE };
+}
+
+/**
+ * Enumerate every `when` predicate in a policy with its evaluation scope and an operator-readable
+ * location. Walks the SAME two locations the runtime evaluates (validators/policy.js):
+ *   - `global_rules[].when` at scope `rule.scope || 'submission'`
+ *   - `surfaces.<surface>.custom_rules[].when` at scope `scenario_result`
+ * `custom_rules` under `defaults` is schema-forbidden, so it is not walked here — the structural
+ * gate already reports it if present (matching the runtime, which never evaluates it).
+ */
+function* collectPredicates(policyDoc) {
+  if (policyDoc === null || typeof policyDoc !== 'object') return;
+
+  const globalRules = Array.isArray(policyDoc.global_rules) ? policyDoc.global_rules : [];
+  for (let i = 0; i < globalRules.length; i++) {
+    const rule = globalRules[i];
+    if (rule && typeof rule === 'object' && rule.when != null) {
+      yield {
+        when: rule.when,
+        scope: rule.scope || 'submission',
+        location: `global_rules[${i}]${rule.id ? ` (${rule.id})` : ''}`,
+      };
+    }
+  }
+
+  const surfaces = policyDoc.surfaces && typeof policyDoc.surfaces === 'object' ? policyDoc.surfaces : {};
+  for (const [surface, sp] of Object.entries(surfaces)) {
+    const customRules = sp && Array.isArray(sp.custom_rules) ? sp.custom_rules : [];
+    for (let i = 0; i < customRules.length; i++) {
+      const rule = customRules[i];
+      if (rule && typeof rule === 'object' && rule.when != null) {
+        yield {
+          when: rule.when,
+          scope: 'scenario_result',
+          location: `surfaces.${surface}.custom_rules[${i}]${rule.id ? ` (${rule.id})` : ''}`,
+        };
+      }
+    }
+  }
+}

package/validators/predicate.js

@@ -134,16 +134,30 @@ function resolvePath(root, segments) {
   return frontier;
 }
 
-/** Validate the leading field segment against the scope's known-field set. */
-function checkLeadingField(field, scope) {
+/**
+ * Decide whether a field's LEADING segment is unknown for a scope, returning the
+ * fault rather than throwing it. Factored out so the eval path ({@link checkLeadingField},
+ * fail-on-first) and the lint path ({@link lintPredicate}, batch-collect) share one
+ * source of truth — a divergence between author-time and runtime diagnostics is
+ * impossible by construction. Data-INDEPENDENT (the known-field set is derived from
+ * the schema), so it is exactly the kind of check the lint can run without a submission.
+ */
+function leadingFieldFault(field, scope) {
   const leading = String(field).split('.')[0].replace('[]', '');
   const known = KNOWN_FIELDS[scope];
   if (known && !known.has(leading)) {
-    throw new PredicateError(
+    return new PredicateError(
       'unknown_field',
       `field "${field}" references unknown leading field "${leading}" (known ${scope} fields: ${[...known].join(', ')})`
     );
   }
+  return null;
+}
+
+/** Validate the leading field segment against the scope's known-field set (eval path: throws). */
+function checkLeadingField(field, scope) {
+  const fault = leadingFieldFault(field, scope);
+  if (fault) throw fault;
 }
 
 /** Apply one operator to a single resolved value + comparand. */
@@ -287,3 +301,149 @@ export function buildReason(rule, root) {
     : (rule.description ?? '');
   return `[${rule.id}] ${body}`;
 }
+
+/* ───────────────────────── Static analysis (VERIFY-F3 policy-lint) ─────────────────────────
+ *
+ * The two functions below analyze a predicate AST WITHOUT a submission. They are the reusable
+ * leaf the `policy-lint` verb (and any future eager-validation path) builds on. Co-located here,
+ * not in a separate module, so they reuse this file's private engine internals (KNOWN_FIELDS via
+ * leadingFieldFault, isCombinator, the depth/node-budget constants) without exporting the guts —
+ * predicate.js owns *everything about a predicate node*: eval AND static analysis.
+ *
+ * Coverage boundary (the VERIFY-F2 over-claim lesson — see docs/policy-lint.md):
+ *   - DATA-INDEPENDENT, caught here: unknown leading field, combinator over-depth, node budget.
+ *   - STRUCTURAL (unknown op, banned/malformed segment, arity): caught earlier by the schema gate
+ *     (`validatePayload('policy', …)`), so lintPredicate does not re-report them.
+ *   - DATA-DEPENDENT, NOT catchable statically: `type_mismatch` (a numeric op over a non-number)
+ *     and `fanout_budget` (a `[]` selection size) — both need a real submission. The lint says so.
+ */
+
+/** A static lint finding: a machine `code` (mirrors {@link PredicateError} codes) + an operator message. */
+
+/**
+ * Statically lint a predicate tree against one scope, batch-collecting the data-independent
+ * semantic faults the schema cannot express. Unlike {@link evaluatePredicate} this NEVER throws —
+ * a lint reports every fault rather than failing on the first — and it is DEFENSIVE: a malformed
+ * node (already flagged by the schema gate) is skipped, not crashed on. The walk mirrors the
+ * evaluator's own depth + node accounting (same constants), so a lint verdict agrees with what the
+ * engine would do at eval time, and the walk is itself bounded (it cannot become a DoS on a
+ * pathological policy file).
+ *
+ * @param {unknown} node - The predicate tree (any value; non-objects are skipped).
+ * @param {'submission'|'scenario_result'} scope - Field-resolution scope.
+ * @returns {{ code: string, field?: string, message: string }[]}
+ */
+export function lintPredicate(node, scope) {
+  const findings = [];
+  walkLint(node, scope, 1, { nodes: 0, depthReported: false, budgetReported: false }, findings);
+  return findings;
+}
+
+function walkLint(node, scope, depth, state, findings) {
+  if (state.budgetReported) return;
+  if (++state.nodes > PREDICATE_MAX_NODES) {
+    if (!state.budgetReported) {
+      state.budgetReported = true;
+      findings.push({
+        code: 'node_budget',
+        message: `predicate exceeds the evaluation budget of ${PREDICATE_MAX_NODES} nodes`,
+      });
+    }
+    return;
+  }
+  if (node === null || typeof node !== 'object') return;
+
+  if (isCombinator(node)) {
+    if (depth > PREDICATE_MAX_DEPTH) {
+      // Mirror the evaluator, which throws when a combinator is reached past the cap; report once
+      // and stop descending this branch (siblings are still walked — batch reporting).
+      if (!state.depthReported) {
+        state.depthReported = true;
+        findings.push({
+          code: 'max_depth',
+          message: `predicate nests deeper than the limit of ${PREDICATE_MAX_DEPTH} combinator levels`,
+        });
+      }
+      return;
+    }
+    const children =
+      Array.isArray(node.all) ? node.all :
+      Array.isArray(node.any) ? node.any :
+      node.not != null ? [node.not] :
+      Array.isArray(node.implies) ? node.implies : [];
+    for (const child of children) walkLint(child, scope, depth + 1, state, findings);
+    return;
+  }
+
+  // Leaf: the only data-independent semantic check is the unknown leading field. The op set and
+  // path shape are schema-gated; a numeric-op type mismatch is data-dependent (not checkable here).
+  if (typeof node.field === 'string') {
+    const fault = leadingFieldFault(node.field, scope);
+    if (fault) findings.push({ code: fault.code, field: node.field, message: fault.message });
+  }
+}
+
+/** Negative operators that fail OPEN over an empty/absent `[]` selection (the footgun). */
+const NEGATIVE_OPS = new Set(['not_equals', 'not_in', 'not_contains', 'not_exists']);
+
+/** The positive counterpart used in the fail-closed `not(any(...))` rewrite suggestion. */
+const POSITIVE_OF = { not_equals: 'equals', not_in: 'in', not_contains: 'contains', not_exists: 'exists' };
+
+/** Build the deterministic, AST-derived fail-closed rewrite suggestion for a footgun leaf. */
+function footgunSuggestion(field, op) {
+  const pos = POSITIVE_OF[op] || 'contains';
+  const valuePart = pos === 'exists' ? ' ' : ', value: … ';
+  return `if you meant "reject when no element satisfies it", use the fail-closed idiom ` +
+    `{ not: { any: [ { field: "${field}", op: ${pos}${valuePart}} ] } }`;
+}
+
+/**
+ * Find `[]`-footgun leaves: a NEGATIVE operator over a `[]` field path, which fails OPEN on an
+ * empty/absent array (existential vacuous truth — see docs/policy-lint.md). ADVISORY only: there
+ * are legitimate existential-negatives, so the caller surfaces these as warnings the author
+ * confirms, never as hard errors, and never auto-applies the rewrite.
+ *
+ * Suppression by negation PARITY, not a bare "is there a `not` above me" flag. A leaf inverted an
+ * EVEN number of times (0, 2, …) still fails open and is flagged; an ODD number of inversions makes
+ * it fail CLOSED, so it is suppressed. Two sources of inversion are counted: a `not` combinator,
+ * and the CONSEQUENT (second element) of an `implies` (since `implies:[A,C]` desugars to the
+ * violation `all(A, not(C))`). This refinement came from the VERIFY-F3 cross-family adversarial
+ * jury (deepseek-v4-pro / glm-5.2 / minimax-m3, 2026-06-30), which converged on two real defects in
+ * the original boolean rule: `not(not(X))` over `[]` fails open but was suppressed (false negative),
+ * and a negative-op consequent of `implies` fails closed but was flagged (false positive). Parity
+ * fixes both. Scope stays NEGATIVE-ops-only and ADVISORY (the jury's "flag every op" and "make it a
+ * hard error" suggestions were rejected — the former floods noise on the normal `contains` idiom,
+ * the latter blocks legitimate existential-negatives). False positives remain acceptable by design;
+ * a false negative (a silent production fail-open) is the expensive failure, so it errs toward warning.
+ *
+ * @param {unknown} node - The predicate tree.
+ * @returns {{ field: string, op: string, suggestion: string }[]}
+ */
+export function findEmptyArrayFootguns(node) {
+  const guns = [];
+  walkFootgun(node, false, { nodes: 0 }, guns);
+  return guns;
+}
+
+function walkFootgun(node, inverted, state, guns) {
+  if (node === null || typeof node !== 'object') return;
+  if (++state.nodes > PREDICATE_MAX_NODES) return; // bounded; lintPredicate reports node_budget
+
+  if (isCombinator(node)) {
+    if (Array.isArray(node.all)) { for (const c of node.all) walkFootgun(c, inverted, state, guns); return; }
+    if (Array.isArray(node.any)) { for (const c of node.any) walkFootgun(c, inverted, state, guns); return; }
+    if (node.not != null) { walkFootgun(node.not, !inverted, state, guns); return; }
+    if (Array.isArray(node.implies)) {
+      // implies:[A, C] === all(A, not(C)) — only the consequent (index 1) flips parity.
+      node.implies.forEach((c, i) => walkFootgun(c, i === 1 ? !inverted : inverted, state, guns));
+      return;
+    }
+    return;
+  }
+
+  if (!inverted &&
+      typeof node.op === 'string' && NEGATIVE_OPS.has(node.op) &&
+      typeof node.field === 'string' && node.field.includes('[]')) {
+    guns.push({ field: node.field, op: node.op, suggestion: footgunSuggestion(node.field, node.op) });
+  }
+}