bareloop 0.8.0 → 0.10.0

package/src/authoring.js

@@ -0,0 +1,1486 @@
+// THE AUTHORING SURFACE — close-authoring v1, module M2 (design record
+// 2026-08-07, FROZEN; gate-2 POC closed 2026-08-08).
+//
+// M1 owns what a stage DOES. This module owns what a declaration may SAY, and
+// whether a given declaration said it legally. Three things live here and they
+// belong together because they are one contract read from three sides:
+//
+//   1. THE KIND CATALOGUE — the entire vocabulary an authoring call has. It
+//      parameterises kinds whose implementations we own; a script body, a shell
+//      fragment and a new kind are not rejected late, they are INEXPRESSIBLE
+//      (D2/D3, the `check-passes(name)` move). `judged-floor` and
+//      `human-confirms` are LOCKED MENU ENTRIES: declaring one is a distinct
+//      `locked-kind` red at VALIDATION, before any tokens, carrying counted
+//      demand (the LOCKED_VERDICTS pattern, src/job.js:520). `harness-loop` is
+//      ABSENT — no entry at all, so declaring it is an unknown-kind typo, which
+//      is the honest reading of a kind that is out of v1 (gate 1, Result 2).
+//
+//   2. THE TYPES GENRE — the one genre v1 admits (D13). Its template is FROZEN
+//      TEXT, byte-identical to the prereg's addendum 3, and its two guard
+//      batteries are OURS. D5 is the load-bearing decision of the whole design:
+//      the close carries guards the user did not ask for and cannot remove,
+//      because the stage that forecloses the easy win is exactly the stage an
+//      LLM asked "what would done look like" will not write well (F87 + the RSI
+//      rubric-close gotcha). Gate-2 round 1 measured the failure mode precisely:
+//      the guard-stripped arm DID invent a guard-SHAPED stage — with 2 of its 4
+//      patterns dead letters and both live suppression channels of that patient
+//      missed. The shape is cheap; the battery is what was paid for.
+//
+//      Round 3 added the two residues addendum 6 hands to this build, and both
+//      are OURS TO INJECT, never model-filled:
+//        - guard SCOPE. The suppression scan covers EVERY changed file of the
+//          genre's extensions. Round 3's arm A narrowed it to the two target
+//          files, which reads stricter and is not: a suppression added in a NEW
+//          helper file walks straight through a target-file filter.
+//        - genre ENV. `MYPYPATH` on a Python patient is the Result-5 fact no
+//          user could supply and no model found (round 2 authored `src`, which
+//          is wrong). It is expressed HERE as genre data and applied by the flow
+//          (M3); a declaration that authors it itself is a red.
+//
+//   3. THE DECLARATION VALIDATOR — the first of D9's three mechanical gates.
+//      Schema, kinds, params, the F84 one-population law, F49's static regex
+//      reject, the D5 guard equality, and hamr's listing rule: every path the
+//      model wrote down must SELECT from the real seed listing. Round 2's arm A
+//      invented `src/alertEmail.js` from the user's prose ("the alert email
+//      one") and silently reclassified 15 real errors into the wrong
+//      population. A path is not a judgment call — it either names something in
+//      the tree or it does not.
+//
+// What this module deliberately does NOT do:
+//   - it never judges QUALITY. Nothing here asks whether the close is a GOOD
+//     close; D9's other two gates (the precheck and the seed-verdict read) plus
+//     the user's signature do that, and no LLM validates another LLM's close.
+//   - it never enforces the genre's stage SKELETON. The template states it to
+//     the author; a close with no work stage is caught by the seed-verdict read
+//     (nothing red at seed = nothing to do), which is where round 2's arm B
+//     died. A validator that demanded named stages would be authoring them.
+//   - it never inspects command ARGS for paths. hamr's rule enumerates path
+//     PARAMS; an arg list carries flags, module names and config paths, and a
+//     `path-not-in-listing` red on `--strict` would be nonsense. The named
+//     consequence, not smoothed: an invented path can still ride in `args`, and
+//     the seed-verdict read is what catches it (it did, in round 2).
+//
+// One spelling of everything shared: `normalizeParser` and `regexGroups` are
+// M1's, imported, never re-spelled — the validator and the executor must agree
+// on what a parser MEANS or the one that grades is the one nobody validated.
+// `hasNestedQuantifier` and `globToPrefix` are the shipped primitives from
+// src/validate.js for the same reason (F49 is monotonic-only doctrine and the
+// detector is never touched from here; globToPrefix's collapse order is fence
+// territory, F9).
+//
+// NOTE for M4 wiring: nothing here is exported from src/index.js yet. M1
+// flagged a `runClose` naming collision between the kind executor and
+// src/ralph.js; the public surface is settled once at M4, not piecemeal.
+
+import { normalizeParser } from './kinds.js';
+import { isObj, isNonEmptyString, hasNestedQuantifier, globToPrefix } from './validate.js';
+
+/** @typedef {{code: string, path: string, detail: string, [k: string]: any}} Red */
+/** @typedef {{required: readonly string[], optional: readonly string[], pathParams: readonly string[], locked?: true, verdictClass: string, shape: string, asserts: string}} KindSpec */
+
+// ── 0. THE VERDICT CLASSES ───────────────────────────────────────────────────
+//
+// The user's own answer (PRD v1.57 §1, restoring the 2026-07-21 radio and
+// superseding D4): is your *done* machine-checkable, does it need judgment, or
+// does it need a person. It is a CLOSED SET handed over enumerated, never a
+// prose rule about what a legal answer would look like.
+//
+// TWO COPIES, PINNED BY THE SUITE RATHER THAN IMPORTED. `src/job.js` owns the
+// same menu as the spec's radio (`VERDICT_TYPES`) and imports this module
+// transitively (job → declaredclose → authoring), so importing it back would
+// close a cycle and hand this module a TDZ-undefined constant at load. The suite
+// asserts the two lists are identical, which is the same treatment
+// `CATALOGUE_LIVE_KINDS` and the executor's `LIVE_KINDS` already get.
+
+/** the whole menu, in ascending order of what it takes to render the verdict */
+export const VERDICT_CLASSES = Object.freeze(['green', 'soft-green', 'hitl']);
+/** declared-but-locked: selecting one is COUNTED demand, never an admission */
+export const LOCKED_CLASSES = Object.freeze(['soft-green', 'hitl']);
+/** the classes v1 actually builds — exactly one */
+export const LIVE_CLASSES = Object.freeze(VERDICT_CLASSES.filter((c) => !LOCKED_CLASSES.includes(c)));
+
+/** the hierarchy as a comparable number. `green` is the FLOOR: a mechanical
+ * measurement is the cheapest honest verdict, a judge is above it, a person is
+ * above that. Read by the ceiling rule and by nothing else. */
+const CLASS_RANK = Object.freeze({ green: 0, 'soft-green': 1, hitl: 2 });
+
+/**
+ * THE CEILING (hamr, 2026-08-08, *"yes to both, go"*): the picked class is a
+ * PROMISE, and this is what the composition actually demands.
+ *
+ * Every kind carries the class of verdict it can honestly render, so a
+ * declaration's ceiling is the HIGHEST class any of its stages needs. A pick
+ * below that ceiling is a silent UPGRADE — a judgment ruler running under a
+ * green promise, which is exactly the fake-hard verdict `CLASS_BY_CLOSE` and
+ * `CLASS_BY_VERDICT` exist to stop, defeated from ABOVE by never letting the
+ * illegal pair be formed.
+ *
+ * INERT IN v1 BY CONSTRUCTION, which is why it is cheap now: every LIVE kind is
+ * mechanical, so no declaration a v1 catalogue can express reaches above green.
+ * It is written before the second class exists so the first soft-green kind is
+ * not the thing that discovers the question.
+ *
+ * A kind the catalogue does not know contributes NOTHING — it already carries an
+ * `unknown-kind` red, and guessing a class for it would be deriving the promise
+ * from a typo. A catalogue entry with no class at all THROWS: the catalogue is
+ * OURS, and failing loudly beats weighing a kind whose honesty nobody stated.
+ * @param {any} declaration
+ * @param {Record<string, KindSpec>} [catalogue]
+ * @returns {{class: string, kind: string, stage: string|null, locked: boolean}|null}
+ */
+export function closeCeiling(declaration, catalogue = KIND_CATALOGUE) {
+  const stages = Array.isArray(declaration?.stages) ? declaration.stages : [];
+  /** @type {{class: string, kind: string, stage: string|null, locked: boolean}|null} */
+  let top = null;
+  for (const s of stages) {
+    if (!isObj(s) || !isNonEmptyString(s.kind)) continue;
+    const spec = catalogue[s.kind];
+    if (!spec) continue;
+    if (!Object.hasOwn(CLASS_RANK, spec.verdictClass)) {
+      throw new Error(`[authoring] catalogue kind "${s.kind}" declares no verdict class — every kind states the class `
+        + `of verdict it can honestly render (one of ${VERDICT_CLASSES.join(', ')})`);
+    }
+    if (top === null || CLASS_RANK[spec.verdictClass] > CLASS_RANK[top.class]) {
+      top = {
+        class: spec.verdictClass,
+        kind: s.kind,
+        stage: isNonEmptyString(s.name) ? s.name : null,
+        locked: spec.locked === true,
+      };
+    }
+  }
+  return top;
+}
+
+// ── 1. THE KIND CATALOGUE ────────────────────────────────────────────────────
+
+/**
+ * The whole vocabulary, as DATA. `shape` and `asserts` are the strings the
+ * authoring prompt renders (M3) — the prompt and this validator must judge the
+ * same facts, so they read one catalogue rather than two (the drift class
+ * SECRET_PATTERNS exists to prevent, one layer up).
+ *
+ * `pathParams` names every param that carries a repository path, in dotted
+ * form. It is data rather than a switch so hamr's listing rule reaches a new
+ * path-bearing param the day one is added, instead of the day someone
+ * remembers to widen a condition.
+ *
+ * `verdictClass` is the class of verdict this kind can HONESTLY render, and it
+ * is what makes the user's pick a promise the composer must keep (PRD v1.57 §1).
+ * Every live kind is `green` — a command's measurement is the truth — so the
+ * ceiling rule is inert in v1 by construction; the two locked kinds carry the
+ * classes they belong to so the rule is already correct the day one goes live.
+ *
+ * Optionality is read off the prereg §5 table, with ONE deliberate divergence
+ * from the gate-2 POC validator, named because it reverses a POC behaviour:
+ * `pattern-absent-in-diff.scope` is OPTIONAL here and the guard ships without
+ * it. Under addendum 6 the guard's scope is GENRE property and the answer is
+ * "every changed file of these extensions" — an absent scope is exactly that
+ * (M1's `inScope` admits everything when no scope object is present). The POC
+ * required it, and required it filled, which is how round 3's arm A came to
+ * narrow the scan and call it stricter.
+ * @type {Record<string, KindSpec>}
+ */
+export const KIND_CATALOGUE = Object.freeze({
+  'command-exit': Object.freeze({
+    verdictClass: 'green',
+    required: Object.freeze(['cmd', 'args', 'expectExit']),
+    optional: Object.freeze(['timeoutMs', 'env']),
+    pathParams: Object.freeze([]),
+    shape: 'cmd: string, args: string[], expectExit: integer, timeoutMs?: integer, env?: {NAME: value}',
+    asserts: 'run cmd/args in the repository; the exit code must equal expectExit.',
+  }),
+  'count-not-worse': Object.freeze({
+    verdictClass: 'green',
+    required: Object.freeze(['cmd', 'args', 'parser', 'scope', 'direction', 'baseline']),
+    optional: Object.freeze(['timeoutMs', 'env']),
+    pathParams: Object.freeze(['scope.includePrefixes', 'scope.excludePrefixes']),
+    shape: 'cmd: string, args: string[], '
+      + 'parser: {terms: [{lineMatch: regex-string, capture?: integer, sign: 1 | -1, '
+      + 'aggregate: "first" | "sum", region: "whole-output" | {anchor: regex-string, capture: integer}}]} '
+      + '— or the short form {lineMatch: regex-string, capture?: integer}, '
+      + 'scope: {includePrefixes?: string[], excludePrefixes?: string[]}, '
+      + 'direction: "lower-is-better" | "higher-is-better", baseline: "seed" | 0, timeoutMs?: integer, env?: {NAME: value}',
+    asserts: 'run cmd/args; read a NUMBER out of the output with parser, keep only what scope selects, and '
+      + 'require that number to be not-worse than baseline in the stated direction. baseline "seed" is the same '
+      + 'number taken from this run\'s own starting point, measured every run; baseline 0 is literally zero. '
+      + 'You never write a number you measured yourself — you write the rule. '
+      + 'If the tool exits NON-ZERO and its output matches none of your parser\'s terms, the stage stops as a broken '
+      + 'instrument rather than recording zero: a crashed checker reporting nothing is unknown, not clean. Name a '
+      + 'tool that prints its count on a clean run.',
+  }),
+  'pattern-absent-in-diff': Object.freeze({
+    verdictClass: 'green',
+    required: Object.freeze(['patterns', 'extensions']),
+    optional: Object.freeze(['scope']),
+    pathParams: Object.freeze(['scope.includePrefixes', 'scope.excludePrefixes']),
+    shape: 'patterns: [{id: string, regex: regex-string}], extensions: string[], '
+      + 'scope?: {includePrefixes?: string[], excludePrefixes?: string[]}',
+    asserts: 'none of the patterns may appear in any line this run ADDED, across every file this run changed '
+      + 'with one of the named extensions.',
+  }),
+  'files-changed': Object.freeze({
+    verdictClass: 'green',
+    required: Object.freeze(['allowPrefixes', 'requireNonEmpty']),
+    optional: Object.freeze([]),
+    pathParams: Object.freeze(['allowPrefixes']),
+    shape: 'allowPrefixes: string[], requireNonEmpty: true',
+    asserts: 'the set of files this run changed must be non-empty and must lie wholly inside allowPrefixes.',
+  }),
+  // LOCKED: named so declaring one is COUNTED DEMAND rather than an unknown-kind
+  // typo (D13 forward-compat point 2). Disclosure is not admission — these do
+  // not run in v1, and `close[]` on disk stays predicate-only and shape-enforced
+  // (src/job.js:442), so the inexpressibility that is already free stays free.
+  'judged-floor': Object.freeze({
+    verdictClass: 'soft-green',
+    required: Object.freeze([]), optional: Object.freeze([]), pathParams: Object.freeze([]), locked: true,
+    shape: 'LOCKED — not available in v1',
+    asserts: 'a judged score must clear a floor. Declaring it is recorded as demand; it will not run.',
+  }),
+  'human-confirms': Object.freeze({
+    verdictClass: 'hitl',
+    required: Object.freeze([]), optional: Object.freeze([]), pathParams: Object.freeze([]), locked: true,
+    shape: 'LOCKED — not available in v1',
+    asserts: 'a person renders the verdict. Declaring it is recorded as demand; it will not run.',
+  }),
+});
+
+/**
+ * Which kinds can carry an environment at all, read off the CATALOGUE rather
+ * than a hand-kept list. ONE spelling, because two consumers must agree
+ * EXACTLY: `applyGenreEnv` injects the genre's variable into these stages and
+ * nowhere else, and the validator's completeness rule demands it on these
+ * stages and nowhere else. A second copy that drifted either way is a hole —
+ * stages injected but unchecked, or stages checked for what nothing injects.
+ * @param {string} kind @param {Record<string, KindSpec>} [catalogue]
+ */
+export function envCapableKind(kind, catalogue = KIND_CATALOGUE) {
+  const spec = catalogue[kind];
+  return !!spec && !spec.locked && [...spec.required, ...spec.optional].includes('env');
+}
+
+export const CATALOGUE_KINDS = Object.freeze(Object.keys(KIND_CATALOGUE));
+export const LOCKED_KINDS = Object.freeze(CATALOGUE_KINDS.filter((k) => KIND_CATALOGUE[k].locked));
+/** the catalogue's live half — asserted EQUAL to the executor's own LIVE_KINDS
+ * by the suite, both directions: a kind the catalogue offers and the executor
+ * cannot run is a close that stops at runtime, and a kind the executor runs
+ * that the catalogue never names is a capability nothing can reach. */
+export const CATALOGUE_LIVE_KINDS = Object.freeze(CATALOGUE_KINDS.filter((k) => !KIND_CATALOGUE[k].locked));
+
+export const DIRECTIONS = Object.freeze(['lower-is-better', 'higher-is-better']);
+/** D12: the close stores the COUNTING RULE, never the number. "seed" is measured
+ * at this run's own seed every run; 0 is literally zero. A frozen constant (the
+ * hand-written closes' `SEED_ERRORS 63`) is exactly what this menu retires. */
+export const BASELINES = Object.freeze(['seed', 0]);
+/** a runaway stage list is a red, not a bill — the sibling of M1's MAX_TERMS */
+export const MAX_STAGES = 12;
+
+const NAME_RE = /^[a-z][a-z0-9-]{1,48}$/;
+
+/**
+ * THE DENY FLOOR under a declared command (hamr's ruling, PRD v1.57 §3: *"there
+ * should be a floor of all dangerous things… straight out block… no one wants
+ * `rm -rf` allowed either ways"*).
+ *
+ * Four families, and each one is here for a different reason:
+ *   - SHELL INTERPRETERS. A shell runs anything, so admitting one makes every
+ *     other entry decoration — the D3 line ("the danger is in the ACTIONS, not
+ *     the SYNTAX") reappearing at the close's own spawn.
+ *   - DESTRUCTIVE / SYSTEM commands. The `rm -rf` class the ruling names, plus
+ *     the machine-level siblings (`dd`, the `mkfs` FAMILY, shutdown/reboot) and
+ *     privilege escalation.
+ *   - FETCH-AND-RUN vectors (`curl`, `wget`). A close that downloads its own
+ *     instrument is a close nobody signed.
+ *   - INDIRECTION (`env`, `eval`, `exec`, `xargs`). They run a program the
+ *     declaration never names, which defeats every entry above by construction.
+ *
+ * It is a FLOOR, and MONOTONIC: names are only ever ADDED, never removed or
+ * narrowed (the F49 precedent — closing a false negative is the allowed
+ * direction; chasing a false positive is not). A legitimate runner refused here
+ * is fixed by declaring the runner, never by shrinking the floor.
+ *
+ * It is NOT A SANDBOX, and nothing here claims to be one. The local-trust model
+ * of v1.41 stands intact and its limitations are written out in PRD v1.57 §3:
+ * close children keep the operator's OS user, network egress and file access;
+ * `args` are not inspected (`node -e <anything>` is reachable and always was);
+ * resource use is bounded by liveness checks, not quotas. The floor narrows
+ * blast radius; it does not promise containment.
+ */
+export const DENIED_COMMANDS = Object.freeze([
+  // shell interpreters
+  'sh', 'bash', 'zsh', 'dash', 'ksh', 'csh', 'tcsh', 'fish',
+  // destructive / system
+  'rm', 'rmdir', 'dd', 'mkfs', 'shutdown', 'reboot', 'halt', 'poweroff',
+  'sudo', 'su', 'chmod', 'chown', 'kill', 'killall', 'pkill',
+  // fetch-and-run
+  'curl', 'wget',
+  // indirection — a program the declaration never named
+  'env', 'eval', 'exec', 'xargs',
+]);
+
+/**
+ * The floor's decision for one declared `cmd`, as a reason or null.
+ *
+ * NORMALISATION FIRST, because that is where the dodge lives: `/usr/bin/rm`
+ * carries the same program as `rm` and only a basename read taken AFTER
+ * splitting the path can see it. Case is folded for the same reason.
+ *
+ * A cmd naming a program OUTSIDE the repository — an absolute path, or one that
+ * climbs out with `..` — is refused whatever its basename says. Two reasons, and
+ * both are the fail-safe direction: the validator has no repository in hand at
+ * every call site (the job validator judges a signed spec with no tree), so
+ * "outside" cannot be decided per-machine; and a machine-absolute path inside a
+ * SIGNED artefact grades a different machine than the one it was signed on.
+ * @param {string} cmd @returns {string|null}
+ */
+export function deniedCommandReason(cmd) {
+  const raw = String(cmd ?? '').trim();
+  if (raw === '') return null; // shape is the schema walk's business, not the floor's
+  const segments = raw.split(/[\\/]/);
+  // `/^[a-zA-Z]:/` is the DRIVE-ABSOLUTE spelling, and it is here because
+  // `scopeContained` (src/validate.js) already counts it as an escape: two
+  // path-escape readings in one codebase that disagree are two instruments, and
+  // this was the one that let `C:\Windows\System32\evil.bat` through to the
+  // basename floor, which had never heard of it. MONOTONIC — this only ever adds
+  // a rejection, which is the direction F49's precedent licenses.
+  if (raw.startsWith('/') || raw.startsWith('\\') || /^[a-zA-Z]:/.test(raw) || segments.includes('..')) {
+    return `"${raw}" names a program outside the repository — a declared close runs the patient's own tools, `
+      + 'reached relatively or through the project\'s runner, never an absolute path (which grades a different '
+      + 'machine than the one that signed it)';
+  }
+  const base = /** @type {string} */ (segments.at(-1)).toLowerCase();
+  const denied = DENIED_COMMANDS.includes(base) || base.startsWith('mkfs.');
+  if (!denied) return null;
+  return `"${raw}" resolves to "${base}", which is on the deny floor — shells, destructive and system commands, `
+    + 'fetch-and-run vectors and program indirection are blocked outright (PRD v1.57 §3). A close names the '
+    + 'patient\'s own instrument; it never names a program that can undo the patient';
+}
+
+// ── 2. THE TYPES GENRE ───────────────────────────────────────────────────────
+
+/**
+ * THE FROZEN TYPES GENRE TEMPLATE — prereg addendum 3's six policy points,
+ * VERBATIM. The suite reads the prereg and compares BYTES, so a paraphrase here
+ * fails the build rather than quietly changing what every future close is
+ * authored against. It deliberately carries no repo-specific command, path, env
+ * value or regex: round 2 measured that stating the LAW still leaves the model
+ * to find the VALUE, and that is the axis the build must keep live.
+ */
+export const TYPES_GENRE_TEMPLATE = `1. The graded instrument is the STRICT form of the language's type checker, regardless of what
+   the repo's own scripts run. If the repo's script omits strictness, the close adds it.
+2. Tools are invoked so their binaries actually resolve — through the project's own package
+   runner or language module runner, never a bare binary name.
+3. The stage skeleton, in first-red-wins order: changed-from-seed (guard) → typecheck (error
+   count IN the target scope, baseline 0, lower-is-better) → typecheck-outside (error count
+   OUTSIDE the target scope, baseline measured at seed, a ceiling — required whenever the job
+   scopes to a subset of the tree; omitted only for whole-tree jobs) → tests-kept (a floor on
+   tests that actually EXECUTED, baseline at seed, higher-is-better; a skipped or deselected
+   test did not run and must not count) → suite-green (the suite exits clean AND reports zero
+   failing tests — two assertions) → no-suppressions (guard).
+4. One population per stage — two structurally different counts never share a stage.
+5. The checker must judge the PATIENT's own tree: if imports could resolve to an installed or
+   editable copy elsewhere, the environment is set so they resolve inside the patient.
+6. A number the tool did not report is unknown, never zero.`;
+
+export const GENRE_LANGUAGES = Object.freeze(['js', 'python']);
+
+/**
+ * THE GENRE, as data. The suppression batteries are the hand-written closes'
+ * own `SUPPRESSIONS` tables (`scripts/u-pulselog-close.mjs:35`,
+ * `scripts/u-spawner-close.mjs:34`) — operator-owned genre knowledge, minted by
+ * paying for live runs. `unknown` is deliberately NOT a JS suppression and not
+ * a Python one: it forces narrowing at the use site, which is real typing work.
+ *
+ * `env` is the addendum-6 residue expressed as data. Python's `MYPYPATH` is
+ * load-bearing and was found by probe, never by reading: the aurora patient sets
+ * `explicit_package_bases`, so without it a sibling import falls through to an
+ * editable install pointing at a DIFFERENT checkout and a cross-module fix is
+ * graded against unedited source. It moves no seed number, which is why no
+ * seed-verdict read can ever catch its absence — the only place it can live is
+ * here, in genre data, injected by the flow.
+ */
+export const TYPES_GENRE = Object.freeze({
+  name: 'TYPES',
+  template: TYPES_GENRE_TEMPLATE,
+  languages: Object.freeze({
+    js: Object.freeze({
+      suppressions: Object.freeze([
+        Object.freeze({ id: 'ts-ignore', regex: '@ts-ignore' }),
+        Object.freeze({ id: 'ts-expect-error', regex: '@ts-expect-error' }),
+        Object.freeze({ id: 'ts-nocheck', regex: '@ts-nocheck' }),
+        Object.freeze({ id: 'eslint-disable', regex: 'eslint-disable' }),
+        Object.freeze({ id: 'any', regex: '(?:[{<|(\\[:,]|\\bas)\\s*any\\b|\\bany\\[\\]' }),
+        Object.freeze({ id: 'any-star', regex: '@\\w+\\s*\\{\\s*[*?]\\s*\\}' }),
+        Object.freeze({ id: 'cast', regex: '@type\\s*\\{.*\\}\\s*\\*\\/\\s*\\(' }),
+      ]),
+      extensions: Object.freeze(['.js', '.mjs', '.cjs']),
+      /** no import-resolution hazard on this runner — an EMPTY list, declared,
+       * never an absent field that would read as "not looked into" (F59) */
+      env: Object.freeze([]),
+      instruments: Object.freeze([
+        Object.freeze({
+          id: 'tsc-error-line',
+          what: 'one type error, as the TypeScript checker prints it',
+          lineMatch: 'error TS\\d+',
+          capture: null,
+          // REAL, from the captured `npm run typecheck -- --strict` of the pulselog
+          // patient (2026-08-09): 67 error lines, every one in this shape.
+          example: 'src/backup.js(21,19): error TS7006: Parameter \'now\' implicitly has an \'any\' type.',
+          why: 'tsc prints a DIFFERENT shape on a terminal (`src/backup.js:21:19 - error TS7006:`) than through a '
+            + 'pipe, and a close only ever reads the piped one. Run msmbpjk6 composed the terminal shape: it matched '
+            + '0 of 67 real error lines while the tool exited 2, so the stage read a crashed instrument. This pattern '
+            + 'is anchored on the CODE, not on the path spelling, so it reads both shapes and cannot lose that flip.',
+        }),
+        // ── node --test, the SAME pipe-vs-terminal trap, MEASURED ──────────────
+        //
+        // PROVENANCE, two independent halves.
+        //
+        // (1) REAL CAPTURED OUTPUT. The mailproof job-#2 battery's own spines carry
+        // whole `npm test` → `node --test` captures inside their `close-precheck`
+        // gap: `bareloop-patients/mailproof-job2-bareloop/battery-P1-mrm7gk25.jsonl`
+        // (`# tests 317` / `# pass 312` / `# fail 5`, five `not ok` lines) and
+        // `battery-P2-mrm7gk25.jsonl` (`# pass 316` / `# fail 1`). Nothing there was
+        // written to match a pattern — it is what the runner printed down the close's
+        // pipe, months before this table existed.
+        //
+        // (2) A SHIPPED, LIVE-FIRED SPELLING. `jobs/pulselog-author-types.json`
+        // already grades `^# pass (\d+)$` and `^# fail (\d+)$` (stages `tests-kept`
+        // and `test-suite-zero-failures`), and run msoc6t8v graded green through it.
+        // The table is not minting a new ruler; it is HANDING OVER the one the tree
+        // already grades with, the way the tsc term does.
+        //
+        // WHY IT IS THE TSC TRAP AGAIN, measured on this machine (node v22.22.2,
+        // one test file, identical args, 2026-08-13): through a PIPE the runner
+        // prints TAP — `TAP version 13`, `# pass 1`, `not ok 2 - b`; on a real pty it
+        // prints the spec reporter — `ℹ pass 1`, no `#`, no TAP at all. A close only
+        // ever reads the piped half, and a pattern composed against the pretty half
+        // matches nothing while the suite runs fine. Exactly what cost run msmbpjk6
+        // its authoring pass on tsc, waiting behind a second tool.
+        Object.freeze({
+          id: 'node-test-total-count',
+          what: 'how many tests the runner EXECUTED, off the TAP summary',
+          lineMatch: '^# tests (\\d+)$',
+          capture: 1,
+          example: '# tests 317',
+          why: 'the EXECUTED count is the honest denominator: a passed-count floor is satisfiable by a suite that '
+            + 'stopped running, so a stage asking "is the suite still whole?" counts this, not passes. Anchored '
+            + 'both ends because the runner prefixes ordinary chatter with the same `# ` (`# Subtest: <name>`), and '
+            + 'an unanchored `# tests` would count test NAMES.',
+        }),
+        Object.freeze({
+          id: 'node-test-pass-count',
+          what: 'how many tests passed, off the TAP summary',
+          lineMatch: '^# pass (\\d+)$',
+          capture: 1,
+          example: '# pass 316',
+          why: 'the spelling `jobs/pulselog-author-types.json` already grades with, green through run msoc6t8v. '
+            + 'The runner prints this line on a green run too, so a zero here is a REPORTED zero.',
+        }),
+        Object.freeze({
+          id: 'node-test-fail-count',
+          what: 'how many tests failed, off the TAP summary',
+          lineMatch: '^# fail (\\d+)$',
+          capture: 1,
+          example: '# fail 1',
+          why: 'this runner prints every counter on every run — a fully green suite still prints `# fail 0` '
+            + '(measured), so reading the figure cannot confuse "no failures" with "never reported". Its pytest '
+            + 'counterpart has no such guarantee, which is why python has no fail-COUNT term.',
+        }),
+        Object.freeze({
+          id: 'node-test-failed-line',
+          what: 'one failing test, as the TAP reporter names it',
+          lineMatch: '^not ok \\d+ ',
+          capture: null,
+          example: 'not ok 7 - edit-renotify: a non-participant edit, and edits on a pending event, ping no one',
+          why: 'one line per failing test, so it TALLIES and no match is a genuine counted zero. The trailing '
+            + 'space and the `^` both carry weight: every passing test prints `ok 7 - …`, and without the anchor '
+            + 'each of those would count as a failure.',
+        }),
+      ]),
+    }),
+    python: Object.freeze({
+      suppressions: Object.freeze([
+        Object.freeze({ id: 'ignore', regex: '#\\s*type:\\s*ignore' }),
+        Object.freeze({ id: 'any', regex: '\\bAny\\b' }),
+        Object.freeze({ id: 'cast', regex: '\\bcast\\s*\\(' }),
+        Object.freeze({ id: 'noqa', regex: '#\\s*noqa' }),
+        Object.freeze({ id: 'mypy-disable', regex: '#\\s*mypy:\\s*disable' }),
+      ]),
+      extensions: Object.freeze(['.py']),
+      env: Object.freeze([
+        Object.freeze({
+          name: 'MYPYPATH',
+          from: 'sourcePrefixes',
+          join: ':',
+          why: 'mypy must name the PATIENT\'s modules: without it a sibling import resolves through an editable '
+            + 'install to a different checkout, and a cross-module fix is graded against unedited source (Result 5, '
+            + 'found by probe). It moves no seed number, so no seed-verdict read can see its absence.',
+        }),
+      ]),
+      instruments: Object.freeze([
+        Object.freeze({
+          id: 'mypy-error-line',
+          what: 'one type error, as mypy prints it',
+          lineMatch: ' error: ',
+          capture: null,
+          // REAL, from a live aurora run's own spine
+          // (bareloop-patients/aurora-u-bareloop/u-ms2c0ls7.jsonl).
+          example: 'packages/spawner/src/aurora_spawner/timeout_policy.py:89: error: Statement is unreachable  [unreachable]',
+          why: 'the operator\'s own hand-written close reads mypy this way (scripts/u-spawner-close.mjs:133), minted '
+            + 'by paid live runs. It is anchored on the SEVERITY word, not on the path spelling, so a relative or '
+            + 'absolute path, a package prefix or a config-driven root all read the same.',
+        }),
+        // ── pytest, from the aurora battery's own captured runs ────────────────
+        //
+        // PROVENANCE: six `close-precheck` gaps in
+        // `bareloop-patients/aurora-soar-bareloop/battery-A{1..4}-*.jsonl` carry the
+        // whole captured pytest run of the aurora patient (pytest 9.1.1, python
+        // 3.14.6, the patient's own `pytest.ini` addopts, which include `-ra`) —
+        // 2,690-odd tests, the `short test summary info` section and the final counts
+        // line. Real, uncrafted, paid for.
+        //
+        // WHAT IS NOT ANCHORED, and why. Every archived counts line arrives padded
+        // with `=` on both sides, and the padding WIDTH is terminal-derived: the
+        // archived long line gets exactly one `=` per side, a short local run gets
+        // twenty-five (measured, pytest 9.0.2, 2026-08-13). Under `-q` the padding is
+        // absent entirely. So the term anchors on the WORD and its number, never on
+        // the decoration — the same discipline as anchoring tsc on the code rather
+        // than the path.
+        Object.freeze({
+          id: 'pytest-pass-count',
+          what: 'how many tests passed, off pytest\'s final counts line',
+          lineMatch: '(\\d+) passed',
+          capture: 1,
+          example: '= 1 failed, 2690 passed, 3 skipped, 18 deselected, 1775 warnings in 326.12s (0:05:26) =',
+          why: 'the ONE segment of this line pytest prints whether the run was green or red, so it is the only '
+            + 'count on it that can be read as a figure without risking "never reported". The other segments '
+            + '(failed, skipped, deselected) are omitted when their count is zero — a green run prints just '
+            + '`1 passed in 0.04s` (measured) — so a term reading THOSE would go silent exactly when the news is '
+            + 'good, and a silent figure term is a broken instrument, not a zero.',
+        }),
+        Object.freeze({
+          id: 'pytest-failed-line',
+          what: 'one failing test, as pytest\'s short summary names it',
+          lineMatch: '^FAILED ',
+          capture: null,
+          example: 'FAILED packages/soar/tests/test_agent_registry_deprecation.py::TestAgentRegistryDeprecation::test_registry_and_discovery_equivalent_results',
+          why: 'this is how you count pytest FAILURES: exactly one line per failing test (measured under both `-ra` '
+            + 'and `-q`), so it TALLIES and no match at all is a legitimate counted zero rather than an unreported '
+            + 'number. It needs pytest\'s short summary section, which the patient\'s own config must not suppress.',
+        }),
+      ]),
+    }),
+  }),
+});
+
+/** @param {string} lang */
+function language(lang) {
+  const l = TYPES_GENRE.languages[/** @type {'js'|'python'} */ (lang)];
+  // An unknown language THROWS rather than returning an empty battery: a genre
+  // we have not admitted must not read as "this genre needs no guards" (F59's
+  // shape — `{}` meaning "the scout did not complete", never "no facts needed").
+  if (!l) throw new Error(`no TYPES genre data for language "${lang}" — one of ${GENRE_LANGUAGES.join(', ')}`);
+  return l;
+}
+
+/**
+ * THE GUARD BATTERY, KEYED BY VERDICT CLASS (hamr's ruling, PRD v1.57 §2).
+ *
+ * D5's battery was genre-keyed data. Under class-keyed interviews it has no
+ * genre to hang from, and the ruling gives it the same home the interview took:
+ * **the battery attaches to the verdict class the user picked** — not to a
+ * problem-genre label, and not per catalogue KIND.
+ *
+ * WHY THE CLASS. What a battery is FOR is the class of dishonesty the class of
+ * VERDICT admits. A green job is a job whose *done* is claimed to be
+ * machine-checkable, so it carries the green battery whatever problem it came
+ * from — that is the anti-suppression obligation the F87 genre earned, and it
+ * must not travel with a label the interview no longer asks for.
+ *
+ * ATTACHMENT POINT vs CONTENTS, and the split is the whole implementation:
+ *   - the battery's STRUCTURE and its un-removability key off the CLASS. That
+ *     is this table: which guards exist, what they assert, which slot is the
+ *     model's (`fill`), and which params are ours.
+ *   - the battery's CONTENTS fill in at COMPOSITION (`compose`), because a
+ *     suppression pattern, a file extension and a checker's escape hatch belong
+ *     to the TOOL the declaration names, and nothing before composition knows
+ *     that tool. `classGuards` resolves them from the language data below.
+ *
+ * A LOCKED class carries `guards: null`, not `[]`. Absent is never empty (F59):
+ * `[]` would read as "this class needs no guards", which is exactly the sentence
+ * D5 exists to make unsayable. Nothing ever reaches it — admission refuses a
+ * locked class before its battery is asked for — and `classGuards` throws rather
+ * than hand back nothing if anything ever does.
+ * @type {Record<string, {locked: boolean, guards: readonly any[]|null}>}
+ */
+export const CLASS_BATTERIES = Object.freeze({
+  green: Object.freeze({
+    locked: false,
+    guards: Object.freeze([
+      Object.freeze({
+        name: 'changed-from-seed',
+        kind: 'files-changed',
+        params: Object.freeze({ requireNonEmpty: true }),
+        compose: Object.freeze([]),
+        fill: Object.freeze(['allowPrefixes']),
+      }),
+      Object.freeze({
+        name: 'no-suppressions',
+        kind: 'pattern-absent-in-diff',
+        // NO scope: the scan covers every changed file of these extensions. A
+        // narrowing here is the round-3 arm A defect and is a red, not a taste.
+        params: Object.freeze({}),
+        compose: Object.freeze(['patterns', 'extensions']),
+        fill: Object.freeze([]),
+      }),
+    ]),
+  }),
+  'soft-green': Object.freeze({ locked: true, guards: null }),
+  hitl: Object.freeze({ locked: true, guards: null }),
+});
+
+/**
+ * How each COMPOSED param resolves from the tool the declaration names. One
+ * entry per composable param name, so a battery that grows a slot fails loudly
+ * here rather than shipping an unresolved one.
+ * @type {Record<string, (l: any) => any>}
+ */
+const COMPOSE_FROM_TOOL = Object.freeze({
+  patterns: (/** @type {any} */ l) => l.suppressions.map((/** @type {any} */ p) => ({ ...p })),
+  extensions: (/** @type {any} */ l) => [...l.extensions],
+});
+
+/**
+ * The picked class's mandatory guards, FULLY ENUMERATED (D13 forward-compat
+ * point 3: an omittable-with-a-default field is the omitted-`tools` shape, where
+ * widening changes what runs without changing what was signed — so the guards
+ * are stored spelled out, and the hash is taken over the resolved form).
+ *
+ * `fill` names the ONLY slot the model supplies. `changed-from-seed`'s
+ * allowPrefixes is the job's own target scope and cannot be ours; everything
+ * else — including `no-suppressions`' absent scope — is.
+ *
+ * AN EMPTY BATTERY IS IMPOSSIBLE for a known class and a known language, and
+ * that is enforced rather than asserted: an unknown class, a locked class, an
+ * unknown language, a battery that resolved to nothing, and a composed param
+ * that resolved to nothing all THROW. F59's shape at the load-bearing decision
+ * of the design — a suppression scan with no patterns reads clean exactly like
+ * one that scanned correctly, and it lands in a SIGNED artefact.
+ *
+ * Returns a fresh deep copy each call: a frozen constant handed to a caller that
+ * edits it is one shared battery away from a silently weakened guard.
+ * @param {{verdictType: string, lang: string}} o
+ * @returns {{name: string, kind: string, params: Record<string, any>, compose: string[], fill: string[]}[]}
+ */
+export function classGuards({ verdictType, lang }) {
+  const battery = Object.hasOwn(CLASS_BATTERIES, String(verdictType)) ? CLASS_BATTERIES[String(verdictType)] : null;
+  if (!battery) {
+    throw new Error(`no guard battery for verdict class "${verdictType}" — one of ${VERDICT_CLASSES.join(', ')}`);
+  }
+  if (battery.locked || battery.guards === null) {
+    throw new Error(`the "${verdictType}" guard battery is LOCKED — that class is declared-but-locked in v1 and `
+      + 'admission refuses it before any battery is built; a locked class must never resolve to an empty one');
+  }
+  const l = language(lang);
+  const guards = battery.guards.map((g) => {
+    /** @type {Record<string, any>} */
+    const params = structuredClone(g.params);
+    for (const p of g.compose) {
+      const resolve = COMPOSE_FROM_TOOL[p];
+      if (!resolve) throw new Error(`the "${g.name}" guard composes "${p}", which nothing knows how to resolve from a tool`);
+      const value = resolve(l);
+      if (!Array.isArray(value) || value.length === 0) {
+        throw new Error(`the "${g.name}" guard's "${p}" resolved EMPTY for language "${lang}" — a guard that checks `
+          + 'nothing reads clean exactly like one that checked correctly, and this one is about to be signed');
+      }
+      params[p] = value;
+    }
+    return { name: g.name, kind: g.kind, params, compose: [...g.compose], fill: [...g.fill] };
+  });
+  if (guards.length === 0) {
+    throw new Error(`the "${verdictType}" battery resolved to NO guards — D5 is the load-bearing decision of this `
+      + 'design and an empty battery is never a legal answer');
+  }
+  return guards;
+}
+
+/**
+ * The genre's environment for this patient — ours to inject at execution (M3),
+ * never model-authored. An entry with no value produces NO VARIABLE: an empty
+ * `MYPYPATH` is not the same instruction as an unset one, and shipping `''`
+ * would silently change how mypy resolves.
+ * @param {string} lang
+ * @param {{sourcePrefixes?: string[]}} [o]
+ * @returns {Record<string, string>}
+ */
+export function genreEnv(lang, { sourcePrefixes = [] } = {}) {
+  const l = language(lang);
+  /** @type {Record<string, string>} */
+  const env = {};
+  for (const spec of l.env) {
+    const parts = (spec.from === 'sourcePrefixes' ? sourcePrefixes : []).filter(isNonEmptyString);
+    if (parts.length) env[spec.name] = parts.join(spec.join);
+  }
+  return env;
+}
+
+/** The names the genre OWNS — the validator refuses a declaration that authors
+ * one of them, because ownership is only real if it is mechanical.
+ * @param {string} lang @returns {string[]} */
+export function genreOwnedEnvNames(lang) {
+  return language(lang).env.map((/** @type {any} */ e) => e.name);
+}
+
+/**
+ * HOW THE GENRE'S OWN TOOLS PRINT — genre property, like the suppression battery
+ * and like `MYPYPATH`, and for the same reason: it is a FACT ABOUT AN INSTRUMENT
+ * WE ALREADY OWN, not a composition choice. Seven hand-written closes read tsc
+ * with the identical `/error TS\d+/` (`scripts/u-pulselog-close.mjs:122`,
+ * `u-bareguard-close.mjs:123`, `u-bareagent-close.mjs:123`,
+ * `u-baremobile-close.mjs:122`, `u-litectx-close.mjs:140`, `types-close.mjs:190`,
+ * `types-check.mjs:60`) and one reads mypy with `' error: '`
+ * (`u-spawner-close.mjs:133`) — the same provenance the SUPPRESSIONS table has:
+ * operator knowledge minted by paying for live runs.
+ *
+ * WHY IT BECAME GENRE DATA (run msmbpjk6, 2026-08-09, $0.43). Left to the model,
+ * the parser for a known tool is a DRAFT-TIME LOTTERY on a fact nobody needed to
+ * guess: that run composed tsc's TTY-only pretty shape
+ * (`^\S+\.js:\d+:\d+ - error TS\d+:`), which matched 0 of the 67 real error lines
+ * the piped run actually printed, while the tool exited 2. The gates held — the
+ * stage read as a crashed instrument and signing was refused, correctly — but the
+ * refusal cost a whole authoring run to learn something the tree already knew. An
+ * earlier run (mslwbkz7) happened to roll the right shape. A coin-flip on an
+ * owned fact is not a capability question.
+ *
+ * WHAT THIS IS NOT: it does not close the parser. A stage counting something we
+ * have NOT measured still writes its own pattern, and that remains a real gap —
+ * named, not smoothed. This table is where a measured instrument fact lands, one
+ * at a time, with its provenance.
+ *
+ * THE TEST-COUNT FACTS (2026-08-13) landed under exactly that rule, and closed
+ * two of the three gaps the paragraph above used to name by hand: the executed
+ * count and the failure line, for both runners, are now measured facts rather
+ * than draft-time lotteries. Provenance rides on each entry.
+ *
+ * WHAT STAYED OUT, deliberately, because no evidence covers it:
+ *  - a pytest FAIL-COUNT figure. A green pytest run omits the segment entirely
+ *    (`1 passed in 0.04s`), so a figure term over it reports "never reported" on
+ *    precisely the runs that went well. Failures are counted with the `^FAILED `
+ *    tally instead, where no match IS zero.
+ *  - a pytest EXECUTED count. `collected N items` is printed under `-ra` and NOT
+ *    under `-q` (measured both ways), so no single spelling survives the modes a
+ *    patient may configure; a stage needing it still composes its own and says so.
+ *
+ * Returns a fresh copy per call, for the reason `classGuards` does: a frozen
+ * constant handed to a caller that edits it is one shared table away from a
+ * silently changed ruler.
+ * @param {string} lang
+ * @returns {{id: string, what: string, lineMatch: string, capture: number|null, example: string, why: string}[]}
+ */
+export function genreInstruments(lang) {
+  const instruments = language(lang).instruments ?? [];
+  // F59's shape at a table that feeds a RULER: an empty list would read as "this
+  // language's tools need no stated format", which is exactly the sentence that
+  // put the pretty-format regex into a signed-close candidate.
+  if (!instruments.length) {
+    throw new Error(`no known-instrument formats for language "${lang}" — a language the genre admits states how its `
+      + 'own checker prints, or the model is back to guessing a fact we already own');
+  }
+  return instruments.map((/** @type {any} */ i) => ({ ...i }));
+}
+
+/**
+ * GROUNDING for a RECORDED genre environment — the provenance half of the
+ * owned-env rule, and the reason the by-value comparison is worth anything.
+ *
+ * The value check proves the envelope's `genreEnv` and the stage's `params.env`
+ * AGREE. Both copies live in the same hand-editable spec, so agreement is
+ * something a forger gets for free: a spec carrying
+ * `MYPYPATH=/tmp/attacker-stubs` in both places is perfectly self-consistent and
+ * would point the checker at a tree nobody signed. Agreement is not provenance.
+ *
+ * What a forger does NOT own is the repository at the seed. Every value the
+ * arbiter can build is `genreEnv`'s join of `sourcePrefixes`, and those prefixes
+ * were already filtered against the seed listing before the join — so each
+ * ELEMENT of a genuine value is a path the listing SELECTS, exactly like hamr's
+ * listing rule for a declared path. Grounding therefore accepts precisely what
+ * `applyGenreEnv` produces and refuses everything else, without needing a second
+ * spelling of what "ours" means.
+ *
+ * Element-wise, never whole-value: a joined list is as strong as its weakest
+ * element, and one real prefix must not launder the rest. A root spelling is
+ * ungrounded here rather than waved through the way `checkPaths` waves it —
+ * `genreEnv` filters those out before joining, so the arbiter can never produce
+ * one, and a validator that admits what its own producer cannot emit is a gap.
+ *
+ * Callers that hold a listing only; the deferred job-spec gate cannot run this
+ * and does not pretend to (see `validateCloseDecl`).
+ * @param {string} lang
+ * @param {Record<string, string>} recorded what the envelope claims the arbiter injected
+ * @param {string[]} listing repo-relative paths at the seed
+ * @returns {{name: string, value: string, element: string, resolved: string}[]} one row per ungrounded element
+ */
+export function ungroundedGenreEnv(lang, recorded, listing) {
+  const idx = indexListing((Array.isArray(listing) ? listing : []).filter(isNonEmptyString));
+  /** @type {{name: string, value: string, element: string, resolved: string}[]} */
+  const out = [];
+  // driven by the GENRE's own specs, not by the recorded keys: the join is genre
+  // knowledge, and a key the genre does not own is the envelope's red, not this
+  // function's business
+  for (const spec of language(lang).env) {
+    const value = recorded?.[spec.name];
+    if (!isNonEmptyString(value)) continue; // shape is the envelope's own check
+    for (const element of String(value).split(spec.join)) {
+      const resolved = normPrefix(element);
+      if (!isRoot(resolved) && (idx.files.has(resolved) || idx.dirs.has(resolved))) continue;
+      out.push({ name: spec.name, value, element, resolved });
+    }
+  }
+  return out;
+}
+
+// ── 3. THE DECLARATION VALIDATOR ─────────────────────────────────────────────
+
+/** one prefix spelling, the SHIPPED one — `globToPrefix`'s collapse order is
+ * fence territory (F9) and is never re-spelled here. `''`, `'.'` and `'/'` all
+ * mean the repository root. */
+const normPrefix = (/** @type {string} */ p) => globToPrefix(String(p)).replace(/^\.\//, '');
+const isRoot = (/** @type {string} */ p) => p === '' || p === '.' || p === '/';
+
+/**
+ * The seed listing, indexed once: the files that exist, and every directory
+ * that contains one. A path SELECTS when it is one of them.
+ * @param {string[]} listing
+ */
+function indexListing(listing) {
+  const files = new Set(listing.map((f) => normPrefix(f)).filter((f) => f !== ''));
+  /** @type {Set<string>} */
+  const dirs = new Set();
+  for (const f of files) {
+    const segs = f.split('/');
+    for (let i = 1; i < segs.length; i++) dirs.add(segs.slice(0, i).join('/'));
+  }
+  return { files, dirs };
+}
+
+/**
+ * Is this a SCOPED job? Mechanically, from the GENRE'S OWN containment guard
+ * against the REAL listing — never from the operator's reading of the interview,
+ * and never from whichever `files-changed` stage happens to come first.
+ *
+ * WHICH STAGE IS READ IS LOAD-BEARING. The candidate set is the genre's guards,
+ * matched BY NAME (the guard battery's own `changed-from-seed`), because that is
+ * the one stage `checkGuards` pins byte-for-byte with `allowPrefixes` as its only
+ * model-filled slot — so its prefixes ARE the job's declared target scope. Read
+ * first-of-kind instead and a declaration carrying a legally-shaped decoy
+ * `files-changed` stage (`allowPrefixes: ['.']`) DECLARED EARLIER reads as
+ * whole-tree, and the F84 one-population law switches itself off silently: the
+ * count stages stop having to say which population they count. Measured, not
+ * imagined — with the decoy the whole declaration validated with zero reds.
+ *
+ * The guard NAMES come from the caller's own `guards` array rather than a literal
+ * here: one spelling, and a genre that renames or adds a containment guard is
+ * covered the day it lands instead of the day someone remembers this function.
+ * No guards in hand means no derivation — `guards-absent` is already a red, and
+ * guessing the job's scope from an unpinned stage is exactly what this fixes.
+ *
+ * A whole-tree job is genuinely unscoped — there is nothing outside — and the
+ * law must not red it (template rule 3 says the outside ceiling is omitted
+ * exactly there).
+ * @param {any} declaration @param {{files: Set<string>, dirs: Set<string>}} idx
+ * @param {any[]|null} guards the picked class's injected guards (`classGuards`)
+ */
+function scopeOfJob(declaration, idx, guards) {
+  const stages = Array.isArray(declaration?.stages) ? declaration.stages : [];
+  const scopeGuards = (Array.isArray(guards) ? guards : [])
+    .filter((g) => isObj(g) && g.kind === 'files-changed' && isNonEmptyString(g.name));
+  for (const g of scopeGuards) {
+    const s = stages.find((/** @type {any} */ x) => isObj(x) && x.name === g.name && x.kind === 'files-changed');
+    if (!s) continue;
+    const pre = s?.params?.allowPrefixes;
+    if (!Array.isArray(pre) || pre.length === 0 || !pre.every(isNonEmptyString)) continue;
+    const norm = pre.map(normPrefix);
+    if (norm.some(isRoot)) continue; // this guard covers the whole tree
+    const uncovered = [...idx.files].some((f) => !norm.some((p) => f === p || f.startsWith(`${p}/`)));
+    if (uncovered) return { scoped: true, via: `${s.name}.allowPrefixes` };
+  }
+  return { scoped: false, via: null };
+}
+
+/**
+ * Kinds a declaration may carry AT MOST ONCE, as data beside the catalogue.
+ *
+ * `files-changed` is the whole list and the reason is the catalogue's own
+ * semantics, not taste. The kind asserts "every changed file lies inside
+ * allowPrefixes"; two of them assert it twice, and stages are ANDed, so the
+ * effective constraint is the INTERSECTION of the two prefix sets — which is
+ * itself a prefix set, and is therefore already expressible as one stage's
+ * `allowPrefixes`. A second containment stage adds no reach whatsoever, so
+ * refusing it costs the author nothing.
+ *
+ * What it does cost to admit is measurable: a second `files-changed` stage is a
+ * second answer to "what is this job's scope", which is the ambiguity
+ * `scopeOfJob` had to be pinned to the guard to survive. This is the same shape
+ * as `count-not-worse`'s `duplicate-population` rule — one population is one
+ * stage — applied to the one population that decides whether the law fires at
+ * all.
+ */
+const AT_MOST_ONCE_KINDS = Object.freeze(['files-changed']);
+
+/** @param {any} v */
+const strArray = (v) => Array.isArray(v) && v.length > 0 && v.every(isNonEmptyString);
+
+/** read a dotted param path off a params object, or undefined @param {any} p @param {string} dotted */
+function readPath(p, dotted) {
+  let node = p;
+  for (const seg of dotted.split('.')) {
+    if (!isObj(node)) return undefined;
+    node = node[seg];
+  }
+  return node;
+}
+
+/**
+ * Validate an authored close DECLARATION. Never throws on model output — every
+ * failure is a named red, in the shipped `{ok, reds, <parsed>}` shape, and the
+ * parsed value is the RESOLVED declaration (every short-form parser expanded),
+ * because that is the form the executor runs and the signature covers.
+ *
+ * `listing`, `guards` and `envOwned` are REQUIRED, and their absence is a red
+ * rather than a skipped check — all three guard the SAME failure: a validator
+ * that silently stops checking when handed nothing to check against reports a
+ * clean declaration it never examined. v1 only admits a job with a runnable
+ * patient and a git seed (D13), so the listing always exists, and D5's guards
+ * exist by construction. `envOwned` is separated into absent-vs-empty for
+ * F59's reason exactly: `[]` is a genre that owns no variables (JS), and
+ * `undefined` is a caller that never asked — the two must not read alike.
+ *
+ * The CATALOGUE is ours, not model input: a malformed one is a bareloop bug and
+ * is allowed to throw, because failing loudly beats a validator that quietly
+ * stops enforcing a rule it was configured out of.
+ *
+ * `deferListing: true` is the ONE way to run without a listing, and it must be
+ * spelled literally. It exists for exactly one caller — the job-spec validator,
+ * which judges a signed `closeDecl` with no repository in hand — and it does NOT
+ * make the listing optional: it moves that half of the gate to the runner, which
+ * re-validates GROUNDED against the real seed before any stage runs. Two checks
+ * ride on the listing and both are named here rather than discovered later: the
+ * path rule (a path SELECTS or it does not exist), and the SCOPED-job derivation
+ * that arms the F84 one-population law. With the listing deferred, neither
+ * fires, and `grounded: false` on the result is how a caller knows.
+ *
+ * `verdictType` is REQUIRED for the same reason: it is the PROMISE the ceiling
+ * rule checks the composition against, and a validator with no pick in hand
+ * cannot check it. Absent or unknown is `class-absent`, naming the closed set.
+ *
+ * @param {any} declaration the parsed model output
+ * @param {{catalogue?: Record<string, KindSpec>, listing?: string[]|null,
+ *   guards?: {name: string, kind: string, params: Record<string, any>, fill: string[]}[]|null,
+ *   envOwned?: string[]|null, envInjected?: Record<string, string>|null, deferListing?: boolean,
+ *   verdictType?: string|null}} [opts]
+ *   `listing` — repo-relative paths at the seed (`git ls-tree -r --name-only`).
+ *   `guards` — the picked class's injected guards (`classGuards`).
+ *   `verdictType` — the class the USER picked (`VERDICT_CLASSES`).
+ *   `envOwned` — env names the genre injects (`genreOwnedEnvNames`), `[]` when it owns none.
+ *   `envInjected` — the env the ARBITER itself injected (`genreEnv`'s output), for a
+ *     POST-injection re-validation. Omitted is the PRE-injection form: every owned name
+ *     is the model's and reds. Supplied, an owned name is accepted only when its value is
+ *     EXACTLY the injected one — a wrong value is the model's guess or corruption.
+ *   `deferListing` — literally `true` to run the tree-independent half only.
+ * @returns {{ok: boolean, reds: Red[], declaration: any, grounded: boolean,
+ *   scoped: {scoped: boolean, via: string|null},
+ *   ceiling: {class: string, kind: string, stage: string|null, locked: boolean}|null}}
+ */
+export function validateDeclaration(declaration, opts = {}) {
+  const {
+    catalogue = KIND_CATALOGUE, listing = null, guards = null, envOwned = null, envInjected = null,
+    verdictType = null,
+  } = opts;
+  const deferListing = opts.deferListing === true;
+  /** @type {Red[]} */
+  const reds = [];
+  /** @type {(code: string, path: string, detail: string, extra?: object) => void} */
+  const red = (code, path, detail, extra = {}) => { reds.push({ code, path, detail, ...extra }); };
+
+  const lockedKinds = Object.keys(catalogue).filter((k) => catalogue[k].locked);
+  const kindNames = Object.keys(catalogue).join(' | ');
+
+  const haveListing = Array.isArray(listing) && listing.length > 0 && listing.every(isNonEmptyString);
+  if (!haveListing && !deferListing) {
+    red('listing-absent', 'listing', 'a non-empty seed file listing (git ls-tree -r --name-only <seed>) — '
+      + 'without it nothing can tell a real path from an invented one, and a validator that skips that check '
+      + 'reports a declaration it never examined');
+  }
+  if (haveListing && deferListing) {
+    red('listing-conflict', 'listing', 'deferListing was declared AND a listing was supplied — a caller that both '
+      + 'defers the tree-grounded half and hands over a tree does not know which gate it is running');
+  }
+  const idx = indexListing(haveListing ? /** @type {string[]} */ (listing) : []);
+
+  const haveGuards = Array.isArray(guards) && guards.length > 0;
+  if (!haveGuards) {
+    red('guards-absent', 'guards', 'the picked class\'s mandatory guards (classGuards) — D5 is the load-bearing '
+      + 'decision of this design and it cannot be validated away by handing the validator nothing to check against');
+  }
+
+  // THE PICK, and it is required for the same reason the guards are: it is the
+  // promise the ceiling rule measures the composition against, and a validator
+  // handed no pick reports a promise it never checked. A closed set, handed over
+  // enumerated — never a prose rule about what a legal class would look like.
+  const picked = VERDICT_CLASSES.includes(String(verdictType)) ? String(verdictType) : null;
+  if (picked === null) {
+    red('class-absent', 'verdictType', `the verdict class the user picked — one of ${VERDICT_CLASSES.join(' | ')}. `
+      + 'It is what a composition promises to stay at or below, and nothing derives it from the declaration');
+  }
+
+  // absent is not empty: `[]` is a genre that owns no environment (JS), and a
+  // caller that never asked is a caller whose declarations are unchecked on the
+  // axis round 2's arm B failed (F59's distinction, in the shape it bites here)
+  const owned = Array.isArray(envOwned) ? envOwned : null;
+  if (owned === null) {
+    red('env-ownership-absent', 'envOwned', 'the env names this genre injects (genreOwnedEnvNames) — pass [] for a '
+      + 'genre that owns none; omitting it leaves a model-authored MYPYPATH-class variable unexamined');
+  }
+  // The arbiter's OWN injection, when this is a POST-injection re-validation.
+  // ABSENT (the default) is the PRE-injection form the authoring loop validates,
+  // where every owned name is the model's and reds — so the loop's gate keeps
+  // exactly the behaviour it had, and only a caller that can state what the
+  // arbiter injected gets to have it accepted.
+  const injected = isObj(envInjected) ? /** @type {Record<string, string>} */ (envInjected) : {};
+
+  if (!isObj(declaration)) {
+    red('invalid-value', '', 'the declaration is an object carrying an ordered stages array');
+    return { ok: false, reds, declaration: null, grounded: haveListing, scoped: { scoped: false, via: null }, ceiling: null };
+  }
+  // NOTES, checked HERE rather than only at the signing gate. The field is the
+  // model's own ("anything you could not express"), it travels verbatim into the
+  // signed `closeDecl`, and `validateCloseDecl` refuses an empty one — so a rule
+  // that lives only there is a rule the authoring loop cannot feed back: the run
+  // pays for a close, accepts it, and dies at signing on a defect the model was
+  // never shown. One spelling of the rule, at the gate that can still revise it.
+  if (declaration.notes !== undefined
+      && !(Array.isArray(declaration.notes) && declaration.notes.every(isNonEmptyString))) {
+    red('invalid-value', 'notes', 'an array of non-empty strings — what the author could not express, kept with '
+      + 'the artefact it is about');
+  }
+
+  const stages = declaration.stages;
+  if (!Array.isArray(stages) || stages.length === 0) {
+    red('missing-field', 'stages', 'a non-empty ORDERED array of stages — they run in the order you declare them');
+    return { ok: false, reds, declaration: null, grounded: haveListing, scoped: { scoped: false, via: null }, ceiling: null };
+  }
+  if (stages.length > MAX_STAGES) {
+    red('invalid-value', 'stages', `${stages.length} stages exceeds the ceiling of ${MAX_STAGES}`);
+  }
+
+  const scoped = scopeOfJob(declaration, idx, guards);
+  // Computed over the WHOLE declaration and REPORTED whatever it says, so a
+  // composition that sits BELOW the pick is visible rather than silent — the
+  // other half of the ruling ("never a silent downgrade"). Only EXCEEDING it is
+  // a red, and that red is raised per offending stage inside the loop.
+  const ceiling = closeCeiling(declaration, catalogue);
+  /** @type {Set<string>} */
+  const seen = new Set();
+  /** @type {Map<string, string>} population key → the stage that already owns it */
+  const populations = new Map();
+  /** @type {Map<string, string>} at-most-once kind → the stage that already owns it */
+  const kindOwners = new Map();
+
+  stages.forEach((/** @type {any} */ s, i) => {
+    const at = `stages[${i}]`;
+    if (!isObj(s)) { red('invalid-value', at, 'a stage is an object {name, kind, params}'); return; }
+
+    if (!isNonEmptyString(s.name)) red('missing-field', `${at}.name`, 'a non-empty stage name');
+    else if (!NAME_RE.test(s.name)) red('invalid-value', `${at}.name`, `"${s.name}" is not a lowercase-hyphenated slug`);
+    else if (seen.has(s.name)) {
+      red('duplicate-name', `${at}.name`, `stage name "${s.name}" is declared twice — two stages under one name are `
+        + 'one trend series over two populations (F84)');
+    } else seen.add(s.name);
+    const label = isNonEmptyString(s.name) ? s.name : at;
+
+    if (!isNonEmptyString(s.kind)) { red('missing-field', `${at}.kind`, `a kind: ${kindNames}`); return; }
+    if (lockedKinds.includes(s.kind)) {
+      // A DISTINCT red carrying counted demand. `verb`/`lib` are the structured
+      // fields the ledger keys admission on, never prose, and `lib` is stamped
+      // HERE at the emit site (the typed-lib rule): a locked KIND is demand
+      // against bareloop's own catalogue, and filing it against a bare-suite
+      // package is the BA-2 misattribution that rule exists to prevent.
+      red('locked-kind', `${at}.kind`, `"${s.kind}" is a named but LOCKED kind — declaring it is recorded demand, `
+        + 'never an admission; it does not run in v1', { kind: s.kind, verb: s.kind, lib: 'bareloop', stage: label });
+      return;
+    }
+    const spec = catalogue[s.kind];
+    if (!spec) {
+      red('unknown-kind', `${at}.kind`, `"${s.kind}" is not in the catalogue — kinds: ${kindNames}`, { kind: s.kind });
+      return;
+    }
+    // THE CEILING (PRD v1.57 §1): composition may only use kinds AT OR BELOW the
+    // picked class. A LOCKED kind never reaches here — it already carries its own
+    // counted-demand red and returned — so a kind refused as unavailable is never
+    // ALSO redded for its class: one problem, one red, in one vocabulary.
+    if (picked !== null && CLASS_RANK[spec.verdictClass] > CLASS_RANK[picked]) {
+      red('class-ceiling', `${at}.kind`, `stage "${label}" uses "${s.kind}", which can only render a `
+        + `${spec.verdictClass} verdict, and this job is declared ${picked}. The class the user picked is a PROMISE: a `
+        + 'composition may use kinds at or below it and never above it, because a judgment ruler running under a '
+        + 'machine-checkable promise is a fake-hard verdict — refused here rather than upgraded silently',
+      { kind: s.kind, stage: label, kindClass: spec.verdictClass, picked });
+    }
+    if (AT_MOST_ONCE_KINDS.includes(s.kind)) {
+      const owner = kindOwners.get(s.kind);
+      if (owner !== undefined) {
+        red('duplicate-kind', `${at}.kind`, `stage "${label}" is a second ${s.kind} stage — "${owner}" already declares `
+          + 'which files this job may change. Stages are ANDed, so two of them mean the intersection of two prefix sets, '
+          + 'which is itself a prefix set and already says the same thing in one stage; what a second one does add is a '
+          + 'second answer to what this job\'s scope IS', { kind: s.kind, stage: label, twin: owner });
+      } else kindOwners.set(s.kind, label);
+    }
+
+    const p = s.params;
+    if (!isObj(p)) { red('missing-field', `${at}.params`, `an object carrying: ${spec.required.join(', ')}`); return; }
+    for (const k of spec.required) {
+      if (!Object.hasOwn(p, k) || p[k] === null || p[k] === undefined) {
+        red('missing-field', `${at}.params.${k}`, `${s.kind} requires ${k} (${spec.shape})`);
+      }
+    }
+    for (const k of Object.keys(p)) {
+      if (!spec.required.includes(k) && !spec.optional.includes(k)) {
+        red('invalid-value', `${at}.params.${k}`, `"${k}" is not a parameter of ${s.kind} — the catalogue is the `
+          + 'only source of parameter names');
+      }
+    }
+
+    checkKind({ kind: s.kind, params: p, at, red, scoped, label, populations, envOwned: owned ?? [], envInjected: injected });
+    checkEnvComplete({ params: p, at, red, kind: s.kind, catalogue, envInjected: injected });
+    checkPaths({ spec, params: p, at, red, idx, haveListing });
+  });
+
+  if (haveGuards) checkGuards({ declaration, guards: /** @type {any[]} */ (guards), red });
+
+  const ok = reds.length === 0;
+  return { ok, reds, declaration: ok ? normalizeDeclaration(declaration) : null, grounded: haveListing, scoped, ceiling };
+}
+
+/**
+ * COMPLETENESS — the deletion half of the owned-env rule, and the half a
+ * by-VALUE comparison structurally cannot cover.
+ *
+ * `genre-owned-env` catches a CHANGED variable: a value that is not the one the
+ * arbiter builds. Nothing caught a DELETED one. Drop the key from a stage — or
+ * drop `params.env` outright — and there is no value left to compare, so a
+ * hand-edited spec validates clean and mypy resolves through an editable install
+ * to a DIFFERENT checkout while the close reports a number. That is the Result-5
+ * hazard exactly, arriving through the one door the value check leaves open, and
+ * it is silent by construction: the variable moves no number the close can see.
+ *
+ * So an injection is a fact about EVERY stage the injector touches, not about
+ * the stages that happen to still carry it. The set is read from the same
+ * catalogue predicate `applyGenreEnv` injects by (`envCapableKind`), so the rule
+ * can never demand a variable on a stage nothing would inject into, nor let one
+ * off a stage that was injected.
+ *
+ * Runs only on a POST-injection re-validation: with nothing recorded there is
+ * nothing to be complete about, and the authoring loop must not ask the model
+ * for the variable it is forbidden to author.
+ * @param {{params: any, at: string, red: (c: string, p: string, d: string, e?: object) => void,
+ *   kind: string, catalogue: Record<string, KindSpec>, envInjected: Record<string, string>}} o
+ */
+function checkEnvComplete({ params: p, at, red, kind, catalogue, envInjected }) {
+  const names = Object.keys(envInjected);
+  if (names.length === 0 || !envCapableKind(kind, catalogue)) return;
+  // a malformed `env` already has its own shape red; saying it twice in two
+  // vocabularies tells the reader there are two problems
+  if (Object.hasOwn(p, 'env') && !isObj(p.env)) return;
+  for (const name of names) {
+    if (isObj(p.env) && Object.hasOwn(p.env, name)) continue;
+    red('genre-env-missing', `${at}.params.env.${name}`, `${name} is recorded as the genre's own injection and this `
+      + 'stage spawns a process, so it must carry it — the arbiter injects into every env-capable stage, and a stage '
+      + 'that comes back without it grades whatever the ambient environment resolves to, silently and with no number '
+      + 'that moves', { name, expected: envInjected[name] });
+  }
+}
+
+/**
+ * The per-kind parameter checks, plus the two halves of the one-population law.
+ * @param {{kind: string, params: any, at: string, red: (c: string, p: string, d: string, e?: object) => void,
+ *   scoped: {scoped: boolean, via: string|null}, label: string, populations: Map<string, string>,
+ *   envOwned: string[], envInjected: Record<string, string>}} o
+ */
+function checkKind({ kind, params: p, at, red, scoped, label, populations, envOwned, envInjected }) {
+  /** F49 — the static nested-quantifier reject, at the gate, before the pattern
+   * ever runs. MEASURED: `(a+)+$` did not finish on a 33-char body in 120s, so
+   * input-bounding is theatre and a static reject is not. The detector itself is
+   * never touched from here (monotonic-only doctrine). */
+  const regex = (/** @type {any} */ v, /** @type {string} */ path) => {
+    if (!isNonEmptyString(v)) { red('invalid-value', path, 'a regular expression, as a string'); return; }
+    if (hasNestedQuantifier(v)) {
+      red('invalid-value', path, `/${v}/ carries a nested quantifier — rejected statically (F49); rewrite it without one`);
+      return;
+    }
+    try { new RegExp(v); } catch (e) {
+      red('invalid-value', path, `/${v}/ is not a valid regular expression: ${String(/** @type {any} */ (e)?.message ?? e)}`);
+    }
+  };
+  const command = () => {
+    if (!isNonEmptyString(p.cmd)) red('invalid-value', `${at}.params.cmd`, 'a non-empty executable name');
+    else {
+      // THE DENY FLOOR (PRD v1.57 §3), at the validation gate, before any token
+      // and before any spawn. A FLOOR, not a sandbox — see DENIED_COMMANDS.
+      const denied = deniedCommandReason(p.cmd);
+      if (denied) red('cmd-denied', `${at}.params.cmd`, denied, { cmd: p.cmd });
+    }
+    if (Object.hasOwn(p, 'args') && !(Array.isArray(p.args) && p.args.every((/** @type {any} */ x) => typeof x === 'string'))) {
+      red('invalid-value', `${at}.params.args`, 'an array of strings (may be empty)');
+    }
+    if (Object.hasOwn(p, 'timeoutMs') && !(Number.isInteger(p.timeoutMs) && p.timeoutMs > 0)) {
+      red('invalid-value', `${at}.params.timeoutMs`, 'a positive integer');
+    }
+    if (Object.hasOwn(p, 'env')) {
+      if (!(isObj(p.env) && Object.values(p.env).every((v) => typeof v === 'string'))) {
+        red('invalid-value', `${at}.params.env`, 'an object of string values');
+      } else {
+        for (const name of Object.keys(p.env)) {
+          if (!envOwned.includes(name)) continue;
+          // The arbiter's OWN injection reaches this check on every re-validation
+          // after the flow applied it (M3's applyGenreEnv), so the split is by
+          // VALUE and never by presence — exactly how `checkGuards` tells a genre
+          // guard from a weakened copy. Presence-flagging here would red the
+          // arbiter's own artefact and no python close could be signed, validated
+          // or run.
+          const injected = Object.hasOwn(envInjected, name);
+          if (injected && p.env[name] === envInjected[name]) continue;
+          // Round 2's arm B authored MYPYPATH itself and authored it WRONG; the
+          // flow sets it from genre data. A value that is not the one the arbiter
+          // builds is the model's guess or a corrupted artefact, either way a red.
+          red('genre-owned-env', `${at}.params.env.${name}`, injected
+            ? `${name} is the genre's own variable and its value is fixed by the genre, not by the declaration: `
+              + `expected ${JSON.stringify(envInjected[name])}, declared ${JSON.stringify(p.env[name])}`
+            : `${name} is set by the genre, not by the declaration — this is a fact the interview cannot supply `
+              + 'and the repository does not state',
+          { name, ...(injected ? { expected: envInjected[name], declared: p.env[name] } : {}) });
+        }
+      }
+    }
+  };
+  /** the one-population law's first half (F84): a count over a scoped job must
+   * say WHICH population it counts, or an in-scope 29 and an outside 4 land in
+   * one trend series and a run that swapped walls reads as converging. */
+  const populationNamed = (/** @type {string} */ path) => {
+    const sc = p.scope;
+    if (!isObj(sc)) { red('missing-field', path, 'a scope object naming this stage\'s population'); return; }
+    const inc = Object.hasOwn(sc, 'includePrefixes') ? sc.includePrefixes : null;
+    const exc = Object.hasOwn(sc, 'excludePrefixes') ? sc.excludePrefixes : null;
+    if (inc !== null && !strArray(inc)) red('invalid-value', `${path}.includePrefixes`, 'a non-empty array of path prefixes');
+    if (exc !== null && !strArray(exc)) red('invalid-value', `${path}.excludePrefixes`, 'a non-empty array of path prefixes');
+    for (const k of Object.keys(sc)) {
+      if (k !== 'includePrefixes' && k !== 'excludePrefixes') red('invalid-value', `${path}.${k}`, 'a scope takes includePrefixes and excludePrefixes only');
+    }
+    if (inc === null && exc === null && scoped.scoped) {
+      red('one-population', path, `stage "${label}" counts over a scoped job (${scoped.via}) with no include/exclude `
+        + 'filter — that mixes the job\'s own population with everything else under one stage name, and the two land '
+        + 'in one trend series (F84)', { rule: 'unnamed-population', stage: label });
+    }
+  };
+
+  switch (kind) {
+    case 'command-exit':
+      command();
+      if (!Number.isInteger(p.expectExit)) red('invalid-value', `${at}.params.expectExit`, 'an integer exit code');
+      break;
+    case 'count-not-worse': {
+      command();
+      // ONE normalisation, the EXECUTOR's own (src/kinds.js). Short form and long
+      // form both land here, so the validator and the grader can never disagree
+      // about what a parser means.
+      const norm = normalizeParser(p.parser, `${at}.params.parser`);
+      for (const r of norm.reds) red(r.code, r.path, r.detail);
+      for (const t of norm.terms ?? []) {
+        regex(t.lineMatch, `${at}.params.parser (lineMatch /${t.lineMatch}/)`);
+        // a region anchor runs over the whole output, so it is exactly as
+        // exposed to F49 as lineMatch is
+        if (isObj(t.region)) regex(t.region.anchor, `${at}.params.parser (region anchor /${t.region.anchor}/)`);
+      }
+      if (!DIRECTIONS.includes(p.direction)) red('invalid-value', `${at}.params.direction`, `one of ${DIRECTIONS.join(' | ')}`);
+      if (!BASELINES.includes(p.baseline)) {
+        red('invalid-value', `${at}.params.baseline`, `one of ${BASELINES.map((b) => JSON.stringify(b)).join(' | ')} — `
+          + 'the close stores the counting rule, never a number you measured (D12)');
+      }
+      populationNamed(`${at}.params.scope`);
+      // the law's second half: the SAME command, parser and scope declared twice
+      // is one population measured twice, wearing two stage names
+      if (norm.ok) {
+        const key = JSON.stringify([p.cmd, p.args ?? [], norm.terms, canonicalScope(p.scope)]);
+        const owner = populations.get(key);
+        if (owner !== undefined) {
+          red('one-population', `${at}.params`, `stage "${label}" measures the same population as "${owner}" — same `
+            + 'command, same parser, same scope. One population is one stage (F84); two names over one number are two '
+            + 'trend series that can never disagree', { rule: 'duplicate-population', stage: label, twin: owner });
+        } else populations.set(key, label);
+      }
+      break;
+    }
+    case 'pattern-absent-in-diff':
+      if (!Array.isArray(p.patterns) || p.patterns.length === 0) {
+        red('invalid-value', `${at}.params.patterns`, 'a non-empty array of {id, regex}');
+      } else {
+        p.patterns.forEach((/** @type {any} */ q, /** @type {number} */ j) => {
+          if (!isObj(q)) { red('invalid-value', `${at}.params.patterns[${j}]`, '{id, regex}'); return; }
+          if (!isNonEmptyString(q.id)) red('invalid-value', `${at}.params.patterns[${j}].id`, 'a non-empty id');
+          regex(q.regex, `${at}.params.patterns[${j}].regex`);
+        });
+      }
+      if (!strArray(p.extensions)) red('invalid-value', `${at}.params.extensions`, 'a non-empty array of file extensions');
+      // scope is OPTIONAL here and absent on the genre guard: the scan covers
+      // every changed file of these extensions. When a declaration does carry
+      // one it is still shape-checked — but a missing scope is not a red.
+      if (Object.hasOwn(p, 'scope')) populationNamed(`${at}.params.scope`);
+      break;
+    case 'files-changed':
+      if (!strArray(p.allowPrefixes)) red('invalid-value', `${at}.params.allowPrefixes`, 'a non-empty array of path prefixes');
+      if (p.requireNonEmpty !== true) red('invalid-value', `${at}.params.requireNonEmpty`, 'literally true');
+      break;
+    default:
+      red('unknown-kind', `${at}.kind`, `"${kind}" is in the catalogue but has no checker`, { kind });
+  }
+}
+
+/** the scope as a comparable value — key order must not decide whether two
+ * stages count the same population @param {any} sc */
+function canonicalScope(sc) {
+  if (!isObj(sc)) return null;
+  const norm = (/** @type {any} */ v) => (Array.isArray(v) ? [...v].map(normPrefix).sort() : null);
+  return { include: norm(sc.includePrefixes), exclude: norm(sc.excludePrefixes) };
+}
+
+/**
+ * HAMR'S LISTING RULE: every path-like param must SELECT from the seed listing.
+ * A path matching nothing is a distinct red, never a judgment call — round 2's
+ * arm A derived `src/alertEmail.js` from the user's prose and reclassified 15
+ * real errors into the wrong population, and no amount of reading the
+ * declaration would have caught it.
+ *
+ * A trailing-slash (or `/**`) spelling ASSERTS a directory, so it must name one;
+ * a bare path may name either. The repository root, however spelled, always
+ * selects — it names the whole tree, which exists.
+ * @param {{spec: KindSpec, params: any, at: string,
+ *   red: (c: string, p: string, d: string, e?: object) => void,
+ *   idx: {files: Set<string>, dirs: Set<string>}, haveListing: boolean}} o
+ */
+function checkPaths({ spec, params: p, at, red, idx, haveListing }) {
+  if (!haveListing) return; // already redded once, at the top; not twice per path
+  for (const dotted of spec.pathParams) {
+    const value = readPath(p, dotted);
+    if (!Array.isArray(value)) continue; // shape is the schema walk's business
+    value.forEach((raw, i) => {
+      if (!isNonEmptyString(raw)) return;
+      const wantsDir = /(?:\/|\/\*\*?)$/.test(raw.trim());
+      const norm = normPrefix(raw);
+      if (isRoot(norm)) return;
+      const path = `${at}.params.${dotted}[${i}]`;
+      const isFile = idx.files.has(norm);
+      const isDir = idx.dirs.has(norm);
+      if (isDir) return;
+      if (isFile && wantsDir) {
+        red('path-not-in-listing', path, `"${raw}" is spelled as a directory but the seed tree says ${norm} is a file`,
+          { declared: raw, resolved: norm, found: 'file', wanted: 'directory' });
+        return;
+      }
+      if (isFile) return;
+      red('path-not-in-listing', path, `"${raw}" matches nothing in the seed tree${nearby(norm, idx)} — a path is `
+        + 'read off the repository, never derived from what a file sounds like it should be called',
+        { declared: raw, resolved: norm, found: null, wanted: wantsDir ? 'directory' : 'any' });
+    });
+  }
+}
+
+/** the mechanical half of the gap: what DOES exist beside the invented path.
+ * Named items, counts, capped — never a paragraph (F38/F39).
+ * @param {string} norm @param {{files: Set<string>, dirs: Set<string>}} idx */
+function nearby(norm, idx) {
+  const parent = norm.includes('/') ? norm.slice(0, norm.lastIndexOf('/')) : '';
+  const under = [...idx.files, ...idx.dirs]
+    .filter((f) => (parent === '' ? !f.includes('/') : f.startsWith(`${parent}/`) && !f.slice(parent.length + 1).includes('/')))
+    .sort();
+  if (under.length === 0) return '';
+  const shown = under.slice(0, 8);
+  const tail = under.length > shown.length ? `, and ${under.length - shown.length} more` : '';
+  return ` (under ${parent === '' ? 'the repository root' : parent} the seed tree has: ${shown.join(', ')}${tail})`;
+}
+
+/**
+ * THE D5 LINE. Every injected guard must be PRESENT, of the right kind, and
+ * byte-equal to the genre's own parameters. Only the guard's declared `fill`
+ * slots are the model's, and a parameter the guard does not carry at all is a
+ * NARROWING — which is how round 3's arm A scoped the suppression scan to two
+ * files and made a weaker guard look stricter.
+ * @param {{declaration: any, guards: any[], red: (c: string, p: string, d: string, e?: object) => void}} o
+ */
+function checkGuards({ declaration, guards, red }) {
+  const stages = Array.isArray(declaration?.stages) ? declaration.stages : [];
+  for (const g of guards) {
+    const i = stages.findIndex((/** @type {any} */ s) => isObj(s) && s.name === g.name);
+    if (i === -1) {
+      red('guard-missing', 'stages', `the genre's "${g.name}" guard is not in the declaration — mandatory guards are `
+        + 'shown to the user and cannot be removed (D5)', { guard: g.name, kind: g.kind });
+      continue;
+    }
+    const s = stages[i];
+    const at = `stages[${i}]`;
+    if (s.kind !== g.kind) {
+      red('guard-weakened', `${at}.kind`, `the "${g.name}" guard is a ${g.kind}; this declares ${String(s.kind)}`,
+        { guard: g.name, param: 'kind' });
+      continue;
+    }
+    const params = isObj(s.params) ? s.params : {};
+    for (const [k, want] of Object.entries(g.params)) {
+      if (!deepEqual(params[k], want)) {
+        red('guard-weakened', `${at}.params.${k}`, `the "${g.name}" guard's ${k} is genre property and is fixed: `
+          + `expected ${JSON.stringify(want)}`, { guard: g.name, param: k, expected: want, declared: params[k] });
+      }
+    }
+    for (const k of Object.keys(params)) {
+      if (Object.hasOwn(g.params, k) || g.fill.includes(k)) continue;
+      red('guard-weakened', `${at}.params.${k}`, `the "${g.name}" guard does not take ${k} — adding one narrows what `
+        + 'the guard covers, and a guard that covers less is not a stricter guard', { guard: g.name, param: k, declared: params[k] });
+    }
+  }
+}
+
+/** structural equality over plain JSON data (the shapes a declaration carries)
+ * @param {any} a @param {any} b */
+function deepEqual(a, b) {
+  if (a === b) return true;
+  if (Array.isArray(a) || Array.isArray(b)) {
+    if (!Array.isArray(a) || !Array.isArray(b) || a.length !== b.length) return false;
+    return a.every((v, i) => deepEqual(v, b[i]));
+  }
+  if (isObj(a) && isObj(b)) {
+    const ka = Object.keys(a);
+    const kb = Object.keys(b);
+    if (ka.length !== kb.length) return false;
+    return ka.every((k) => Object.hasOwn(b, k) && deepEqual(a[k], b[k]));
+  }
+  return false;
+}
+
+/**
+ * The declaration the EXECUTOR sees and the SIGNATURE covers: every
+ * `count-not-worse` parser resolved to `{terms: [...]}`, so the short form stops
+ * existing at this boundary and there is exactly one shape downstream (D13
+ * forward-compat point 3 — the hash is taken over the RESOLVED form, or an
+ * omittable-with-a-default field changes what runs without changing what was
+ * signed).
+ *
+ * A parser that would not normalise is left VERBATIM: it already carries a red,
+ * and rewriting it would hide what the model actually wrote.
+ * @param {any} declaration
+ */
+export function normalizeDeclaration(declaration) {
+  if (!isObj(declaration) || !Array.isArray(declaration.stages)) return declaration;
+  const out = structuredClone(declaration);
+  for (const s of out.stages) {
+    if (!isObj(s) || s.kind !== 'count-not-worse' || !isObj(s.params)) continue;
+    const n = normalizeParser(s.params.parser, 'parser');
+    if (n.ok) s.params.parser = { terms: n.terms };
+  }
+  return out;
+}