@dzhechkov/harness-cli 0.8.24 → 0.8.26

package/src/cli.ts

@@ -4,8 +4,16 @@
  * @packageDocumentation
  */
 
-import { appendFileSync, chmodSync, closeSync, cpSync, existsSync, fstatSync, lstatSync, mkdirSync, mkdtempSync, openSync, readFileSync, readSync, readdirSync, readlinkSync, realpathSync, renameSync, rmdirSync, rmSync, statSync, symlinkSync, unlinkSync, writeFileSync, type Dirent } from 'node:fs';
-import { basename, dirname, extname, isAbsolute, join, relative, resolve, sep } from 'node:path';
+import { parseNpmPackInventory, type InventorySource, type LocalInventoryResult } from '@dzhechkov/harness-core';
+// Fix-round 1 (Codex HIGH-1c, feature recall-short-terms): the ONE place `dz recall` prints an
+// empty result must name WHY — via the shared helper, not by re-deriving the decision. Routed
+// through harness-core's re-export (lead correction) rather than a new direct dependency on
+// `@dzhechkov/memory`: a new package-graph edge is a publishing-surface change outside this
+// feature's scope, and harness-core already depends on memory.
+import { noSearchableTermsReason } from '@dzhechkov/harness-core';
+import { appendFileSync, chmodSync, closeSync, constants as fsConstants, cpSync, existsSync, fstatSync, fsyncSync, lstatSync, mkdirSync, mkdtempSync, openSync, readFileSync, readSync, readdirSync, readlinkSync, realpathSync, renameSync, rmdirSync, rmSync, statSync, symlinkSync, unlinkSync, writeFileSync, writeSync, type Dirent } from 'node:fs';
+import { basename, dirname, extname, isAbsolute, join, relative, resolve, sep, posix as nodePosixPath } from 'node:path';
+const posixNormalize = nodePosixPath.normalize;
 import { fileURLToPath } from 'node:url';
 import { request as httpsRequest } from 'node:https';
 import { KNOWN_CLI_FLAGS } from './known-flags.js';
@@ -14,7 +22,7 @@ import { resolveInstallSpec } from './install-spec.js';
 import { dispatchedCommands, documentedCommands } from './command-inventory.js';
 import { execFile, execFileSync, execSync, spawn, spawnSync, type ChildProcess } from 'node:child_process';
 import { createHash, randomBytes } from 'node:crypto';
-import { homedir, hostname, tmpdir } from 'node:os';
+import { cpus, homedir, hostname, tmpdir } from 'node:os';
 import { createRequire } from 'node:module';
 import { isDeepStrictEqual } from 'node:util';
 import { JOURNAL_KINDS, formatLine, parseLine, selectWindow, appendWitnessed, type JournalKind, type JournalIo } from '@dzhechkov/harness-core';
@@ -64,6 +72,7 @@ import {
   runDoctor,
   runInit,
   discoverSkillIds,
+  loadSkillFromDir,
   resolveSelection,
   formatSelectRefusal,
   runIntegrationsVerify,
@@ -138,6 +147,7 @@ import {
   TARGET_SHORT_LABELS,
   applyLegStatus,
   applyLegReasonMessage,
+  probeApplyLeg,
   resolveAgentdbPath,
   WORKFLOW_TEMPLATES_RETIRED_MESSAGE,
   parsePlan,
@@ -557,6 +567,7 @@ import {
   REGISTRY_SELFCHECK_TESTS,
   buildMutationTestCommand,
   parseMutationRegistry,
+  registryEntriesAddedSince,
   applyMutationToText,
   attributeBaselineRedness,
   countFailingTests,
@@ -629,7 +640,7 @@ import {
 import type { OperatorProfile, Domain as ProfileDomain, Register as ProfileRegister } from '@dzhechkov/harness-core';
 import type { RecordKind, RecordDecision } from '@dzhechkov/harness-core';
 import type { CadenceWindow } from '@dzhechkov/harness-core';
-import type { MutationEntryResult, MutationObservation, MutationRegistryEntry } from '@dzhechkov/harness-core';
+import type { MutationEntryResult, MutationObservation, MutationRegistryEntry, MutationRegistry } from '@dzhechkov/harness-core';
 import type { SkillApplyFailure, SkillLoadFailure } from '@dzhechkov/harness-core';
 import type { ReqeDebt } from '@dzhechkov/harness-core';
 import type { ClassifyResultRow, ExecutionEvidence } from '@dzhechkov/harness-core';
@@ -751,7 +762,7 @@ Usage:
   dz profile [init|show|set|sync] [--json]   (WHO the assistant is talking to — per-user store at ~/.dz/profile.json (0600, NEVER in a project), delivered as a marked block in ~/.claude/CLAUDE.md so it loads in EVERY project, dz installed or not. init = five questions (language, register, deep/weak domains as comma lists — "networking (CCIE; NSX)" keeps the parenthetical as the note, Enter skips — teaches y/n with one re-ask, never a silent default); show ALWAYS prints the store path + age + drift verdict + the rendered block; set register|language|teaches <v> or set deep|weak add|rm <tag> [note] — register accepts the owner's own words (профи / профи лайт / просто), an unknown value is REFUSED naming the accepted set; sync re-writes the block (runs automatically after init/set; foreign content byte-for-byte, timestamped backup before every modifying write). The register changes FORM, never FACTS, and governs dialogue only — never ADRs/commits/QE reports; both rules are baked into the rendered block at every level. exit 0 done / 1 no profile or failed / 2 refused input)
   dz reqe [--slug <feature> [--done --report <f>]] [--json]   (the re-QE debt ledger: a usage-switched run whose Step-8 QE ran on the coder's OWN family records a debt; list debts, print the cross-family review brief, settle FAIL-CLOSED against a graded report — the settlement lands in 08_qe_report.md)
   dz qe-bridge --family claude --slug <feature> [--coder-family codex|claude] [--model <id>] [--files a,b] [--out <f>] [--timeout <s>] [--allow-same-family] [--json]   (the REVERSE QE bridge: run an INDEPENDENT Claude reviewer over a feature's Step-8 artifacts from ANY host — a Codex session included, plain shell, no Claude agent plane needed — and land a PARSED signoff. The reviewer runs ISOLATED: an EMPTY temp cwd plus --safe-mode --strict-mcp-config --tools '' --no-session-persistence, so no CLAUDE.md/skills/plugins/hooks/MCP load, and the verdict is read from the --output-format json RESULT ENVELOPE — text a session customization printed onto the same stdout can never become a signoff. Probes the model before trusting it; sends SCOPED extracts with a loud 200k-char ceiling (never silent truncation); the grade must AGREE across three LAST-anchored channels (terminal marker line, fenced qe-bridge-signoff JSON, the report's own GRADE line) AND the marker must be the FINAL content — empty, gradeless, self-contradicting or miscounted output is one of 17 NAMED failures with an audit record under features/<slug>/.fa-state/qe-bridge/ (runId, resolved executable + binOverride, prompt sha256, channel offsets, requestedOut, reportWritten, retained raw stdout; 0600 files in a 0700 dir), never a clean review. A --coder-family that contradicts the recorded reqe debt is refused. Writes features/<slug>/08b_reqe_report.md, which dz reqe --done settles unchanged. DISCLOSURE: the extracts you scope are sent to the Claude runtime; the bridge cannot classify secrets. DZ_QE_BRIDGE_CLAUDE_BIN is a TEST SEAM, not a flag. exit 0 signoff parsed (ANY grade — it reports, it does not gate) / 1 named failure / 2 usage)
-  dz mutation-gate [--package <dir>] [--registry <file>] [--test-cmd "<cmd>"] [--only <id[,id]>] [--timeout <ms>] [--rebaseline per-entry|final] [--keep-scratch] [--json]   (prove each NAMED protection has a test that DISCRIMINATES: copy the package to a scratch dir, verify the baseline suite is green, apply each registry mutation, run the suite, REQUIRE red, restore. The red must be BEHAVIOURAL: a mutation that no longer parses is MUTATION_UNPARSEABLE; a red run whose OWN output reports a test FILE failing to load (node --test file-level not-ok with exitCode, vitest Failed Suites) is MUTATION_LOAD_FATAL — the signal comes from the same run as the failing count, never from a separate isolated import; red output whose shape matches no known runner is INCONCLUSIVE (a runner-coverage gap, loud, never PROVEN); a count far above the entry's bound is OVER_FAILING; a restored tree that does not reproduce green makes the entry INCONCLUSIVE (flaky). Mutation writes are realpath-contained to the scratch copy: a symlink escape or a node_modules/ target is refused (exit 2), the real tree is never written. A mutation that does not apply, a green suite, or an inconclusive run is a FAILURE — never a skip. exit 0 all proven / 1 gate failed / 2 setup error)
+  dz mutation-gate [--package <dir>] [--registry <file>] [--test-cmd "<cmd>"] [--only <id[,id]>] [--touched <path[,path]>] [--added-since <git-ref>] [--timeout <ms>] [--max-workers <n>] [--rebaseline per-entry|final] [--keep-scratch] [--json]   (prove each NAMED protection has a test that DISCRIMINATES: '--max-workers' resolves flag > the registry's own 'maxWorkers' field > 'min(4, max(1, floor(cpus/2)))' (an invalid flag value — 0, negative, fractional, non-numeric — is a usage error, exit 2, never a silent default; fix-round 1), injects '--maxWorkers=<n>' right after 'vitest run' inside its own compound-command segment (unless that segment already names the flag; fix-round 1 — scoped detection, not a whole-command substring check) and sets 'VITEST_MAX_WORKERS=<n>' in the env regardless — an uncapped full-suite baseline/mutant run at vitest's default worker count (= cpu cores) has measured load 62-358 and <2GB free on an 8-core/16GB box under embedding-daemon tests, killing full overnight gate runs (0bb74d66); printed as 'mutation-gate: workers: <n> (<flag|registry|default>)', or 'mutation-gate: workers: n/a — test command is not vitest' when the command is not recognised as vitest. '--touched' selects entries whose 'file' matches one of the given paths, accepted in ANY of package-relative, './'-prefixed, absolute-inside-the-package, repo-relative, or backslash-separated form — all normalized to package-relative POSIX before matching (fix-round 1, AM-1); a path that resolves OUTSIDE the package is counted, never silently dropped, as '<K> outside package' in the 'selected N of M' line; '--added-since <ref>' selects entries whose id is not present in the registry as it read at that ref ('git show <ref>:<registry path>'): a registry genuinely ABSENT at that ref means every current entry counts as added (said explicitly); an unresolvable ref, any OTHER git failure, or an invalid/malformed base registry at that ref is a usage error (exit 2), never folded into "absent" (AM-3) — a feature scopes the gate to its own touched files and any entries it just added instead of the whole registry (MEASURED: an unscoped run over 358 entries on this repo's core package ran 30-40 minutes and hit the timeout wall, INCONCLUSIVE every time). The two selectors UNION and the result INTERSECTS with '--only' when both are given; an empty selection prints 'selected 0 of M entries (…)' and exits 0 — never a silent skip; '--json' always carries a 'selection' object ({selected, total, touched, addedSince, base, outsidePackage}) on every scoped run (AM-4). copy the package to a scratch dir, verify the baseline suite is green, apply each registry mutation, run the suite, REQUIRE red, restore. The red must be BEHAVIOURAL: a mutation that no longer parses is MUTATION_UNPARSEABLE; a red run whose OWN output reports a test FILE failing to load (node --test file-level not-ok with exitCode, vitest Failed Suites) is MUTATION_LOAD_FATAL — the signal comes from the same run as the failing count, never from a separate isolated import; red output whose shape matches no known runner is INCONCLUSIVE (a runner-coverage gap, loud, never PROVEN); a count far above the entry's bound is OVER_FAILING; a restored tree that does not reproduce green makes the entry INCONCLUSIVE (flaky). Mutation writes are realpath-contained to the scratch copy: a symlink escape or a node_modules/ target is refused (exit 2), the real tree is never written. A mutation that does not apply, a green suite, or an inconclusive run is a FAILURE — never a skip. exit 0 all proven / 1 gate failed / 2 setup error)
   dz backlog add "<idea>" [--effort 1-5] [--proposal <text>] [--dry-run] [--allow-cold-start] [--project <dir>] [--json]   (capture an idea: semantic dedup against existing ideas via the Brain vector engine (DUPLICATE>=0.92 merges, RELATED links, NEW creates) + GoalMap alignment; --dry-run classifies without writing)
   dz backlog list [--status <s>] [--goal <id>] [--project <dir>] [--json]   (list captured ideas, filterable by status/goal)
   dz backlog show <id> [--project <dir>] [--json]                          (full record for one idea)
@@ -854,6 +865,13 @@ export type MutationGateRunner = (
     readonly timeoutMs: number;
     readonly phase: 'baseline' | 'mutation' | 'rebaseline' | 'final-rebaseline';
     readonly entryId?: string;
+    /**
+     * mutation-gate-baseline-honesty FR-3: the extra env vars the REAL executor would set on top of
+     * `process.env` for this run — currently just `VITEST_MAX_WORKERS`, set unconditionally
+     * regardless of whether the test command is recognised as vitest (a non-vitest command still
+     * gets the env var; only the command-string injection is vitest-gated).
+     */
+    readonly env: Readonly<Record<string, string>>;
   },
 ) => MutationGateRunnerObservation;
 
@@ -955,6 +973,20 @@ export interface CliIo {
     command: string,
     options: { cwd?: string | URL | undefined; stdio?: unknown; encoding?: unknown; timeout?: number | undefined; env?: NodeJS.ProcessEnv | undefined },
   ) => string;
+  /**
+   * Test seam for `dz publish`'s gate-audit writer (feature `publish-gate-audit-durable`, FR-2):
+   * overrides the fs primitives `appendPublishGateAudit` uses for its durable append (production
+   * leaves it unset → the real `node:fs` functions). Lets a test make `fsyncSync` throw to prove
+   * `(audit NOT logged: …)` is printed and the write is reported as failed, without touching any
+   * other seam's filesystem.
+   */
+  readonly publishGateAuditFsLayer?: PublishGateAuditFsLayer;
+  /**
+   * AM-5 (feature publish-gate-audit-durable): test seam for the sibling-drift gate's `npm pack
+   * --dry-run --json` call (production leaves it unset → real `execFileSync`). Takes the package
+   * dir, returns raw stdout, or throws to simulate a real `npm` failure without spawning anything.
+   */
+  readonly publishNpmPackRunner?: (dir: string) => string;
   /**
    * Test seam for `dz install`: overrides the `npm install` subprocess (production leaves
    * it unset → real `execSync`, stdio piped). A stub runner that pre-stages a fixture
@@ -2584,6 +2616,41 @@ async function cmdInstall(
   if (root.layout === 'npx-template' && root.hasCompanionAssets) {
     write(`  note: ${pkg} also ships commands/hooks/agents — \`npx -y ${pkg} init\` installs the full kit.`);
   }
+  // Junk-skip summary (feature skills-walk-symlinks-and-junk, FR-2/AC-3). `report.skills[].skipped`
+  // above is a WRITE outcome (an existing file not overwritten without --force) — a different
+  // question from "did this skill directory contain build/cache junk that never became an asset at
+  // all". Re-walk each discovered skill (a cheap second READ — `loadSkillFromDir` already did this
+  // once inside `runInit`'s `adapter.compile`, this adds no write) to surface that count without
+  // reshaping `InitReport`. Silent at N=0 (AC-3): most packages ship no junk and must print nothing.
+  // fix-round 1 MEDIUM-4: this counts ENTRIES, not files. A skipped junk DIRECTORY (e.g.
+  // `__pycache__`) is exactly one entry here even though it may hold many files underneath —
+  // `walkFiles` never descends into a skipped junk directory to count those (see skills.ts),
+  // so a per-file count would be a number this code cannot honestly produce. The wording and
+  // the trailing `/` on directory paths say so, instead of implying "file" for something that
+  // may be a whole tree.
+  let junkSkippedCount = 0;
+  const junkSkippedPaths: string[] = [];
+  for (const id of discoverSkillIds(root.dir)) {
+    let skill;
+    try {
+      skill = loadSkillFromDir(root.dir, id);
+    } catch {
+      continue; // unparseable skills are reported separately via report.failures below
+    }
+    for (const entry of skill.skipped ?? []) {
+      if (!entry.reason.startsWith('junk')) continue; // broken-symlink/cycle/escape skips are a different concern (FR-1/AM-8)
+      junkSkippedCount += 1;
+      const shownPath = relative(pkgDir, entry.path).split('\\').join('/');
+      const isDir = entry.reason.startsWith('junk directory');
+      junkSkippedPaths.push(isDir ? `${shownPath}/` : shownPath);
+    }
+  }
+  if (junkSkippedCount > 0) {
+    const shown = junkSkippedPaths.slice(0, 5).join(', ');
+    const more = junkSkippedPaths.length > 5 ? `, … (+${junkSkippedPaths.length - 5} more)` : '';
+    const noun = junkSkippedCount === 1 ? 'entry' : 'entries';
+    write(`skills: skipped ${junkSkippedCount} junk ${noun} (${shown}${more})`);
+  }
   // Skip-and-collect at install time (D1 / the report's D2 amendment): the offending
   // SKILL.md came out of the DOWNLOADED TARBALL, so the path is rendered relative to
   // the package root (a `node_modules/**` absolute path is not actionable) and the
@@ -5046,7 +5113,18 @@ async function cmdRecall(
 
   if (hits.length === 0) {
     write(`dz recall "${shownQuery}"`);
-    write(`  No matching patterns (teach some with \`dz teach\`, or consolidate sessions).`);
+    // Fix-round 1 (Codex HIGH-1c): a query that tokenizes to literally nothing (pure
+    // punctuation/whitespace) is a DIFFERENT empty result than a query with real terms that
+    // simply matched no record — FR-3 requires the reason to be named, not folded into the same
+    // generic "no matching patterns" line. `noSearchableTermsReason` is the single source of
+    // truth for the decision (memory/src/tokenize.ts); this print site calls it rather than
+    // re-deriving "no searchable terms" from the query text itself.
+    const noTermsReason = noSearchableTermsReason(query);
+    if (noTermsReason !== undefined) {
+      write(`  no searchable terms in "${shownQuery}" (only punctuation/whitespace) — reason: ${noTermsReason}`);
+    } else {
+      write(`  No matching patterns (teach some with \`dz teach\`, or consolidate sessions).`);
+    }
     // The domain note must print here too (Codex QE #10): a --domain run with zero hits
     // silently said nothing about the domain, so the reader could not tell whether the
     // boost had been applied and found nothing, or had not run at all.
@@ -7035,6 +7113,77 @@ function packedInstallScratchRoot(): string {
   return existsSync('/var/tmp') ? '/var/tmp' : tmpdir();
 }
 
+/**
+ * FR-1 (feature release-smoke-staged-pack): stage every target's `package.json` exactly like a
+ * live publish packs it — `workspace:*` sibling specs rewritten to the exact sibling version
+ * (`rewriteWorkspaceSpecs`), `scripts.prepublishOnly` dropped — run `fn`, then ALWAYS restore the
+ * original bytes in a `finally`, whatever `fn` does or throws. A restore failure is reported
+ * through `write` (with the path), never swallowed — the "absence of a receipt is not success"
+ * rule this file follows everywhere else.
+ *
+ * `cmdPublish`'s dry-run preview and `cmdRelease`'s packed-install-smoke `pack` steps both go
+ * through this ONE helper, so the two doors that ask "what would the registry receive?" pack the
+ * exact same bytes (MEASURED 2026-09-13 16:05: `dz release` packed the live `workspace:*`
+ * package.json and its smoke install died with EUNSUPPORTEDPROTOCOL — `dz publish`'s preview
+ * already staged around this and release never got that).
+ */
+function withStagedPackageJson<T>(
+  targets: readonly { readonly dir: string }[],
+  workspaceVersions: ReadonlyMap<string, string>,
+  write: Write,
+  fn: () => T,
+  label = 'dz',
+): T {
+  // Lead edits after Codex review (2026-09-13, findings 1/5/6): every write — the staged text and
+  // the restore — goes through a sibling temp file + rename, so a reader never sees a truncated
+  // package.json; a restore that FAILS is an error the caller must see (thrown after fn, or attached
+  // to fn's own error), never a warning that lets a run "succeed" on a damaged tree; and the
+  // diagnostic keeps the calling command's name (`label`).
+  const atomicWrite = (path: string, text: string): void => {
+    // Codex round 2: an EXCLUSIVE, randomized sibling temp — never a shared pid-named file
+    const tmp = `${path}.${process.pid}.${randomBytes(6).toString('hex')}.staged.tmp`;
+    writeFileSync(tmp, text, { flag: 'wx' });
+    renameSync(tmp, path);
+  };
+  const stagedOriginals: Array<{ path: string; text: string }> = [];
+  let fnError: unknown;
+  let fnThrew = false;
+  try {
+    for (const p of targets) {
+      const pkgJsonPath = join(p.dir, 'package.json');
+      const original = readFileSync(pkgJsonPath, 'utf-8');
+      const rewritten = JSON.parse(rewriteWorkspaceSpecs(original, workspaceVersions)) as Record<string, unknown>;
+      const scripts = rewritten['scripts'];
+      if (scripts !== null && typeof scripts === 'object' && !Array.isArray(scripts)) delete (scripts as Record<string, unknown>)['prepublishOnly'];
+      stagedOriginals.push({ path: pkgJsonPath, text: original });
+      atomicWrite(pkgJsonPath, JSON.stringify(rewritten, null, 2) + '\n');
+    }
+    return fn();
+  } catch (err) {
+    fnThrew = true;
+    fnError = err;
+    throw err;
+  } finally {
+    const restoreFailures: string[] = [];
+    for (const o of stagedOriginals) {
+      try {
+        atomicWrite(o.path, o.text);
+      } catch (err) {
+        const msg = `${label}: ✗ could not restore ${o.path} after staged packing: ${formatPublishError(err)} — the tree is left STAGED, restore it by hand`;
+        try { write(msg); } catch { /* a throwing writer must not mask the restore failure */ }
+        restoreFailures.push(msg);
+      }
+    }
+    if (restoreFailures.length > 0) {
+      // Codex round 2: one aggregate error carrying BOTH fn's own failure (if any) and the restore
+      // failures — never a bare message assignment that could itself throw out of finally.
+      const fnPart = fnThrew ? `\n(during: ${fnError instanceof Error ? fnError.message : String(fnError)})` : '';
+      // eslint-disable-next-line no-unsafe-finally -- a damaged tree must not read as success
+      throw new Error(`${restoreFailures.join('\n')}${fnPart}`);
+    }
+  }
+}
+
 function cmdPublish(
   options: Map<string, string>,
   flags: Set<string>,
@@ -7047,6 +7196,14 @@ function cmdPublish(
     command: string,
     options: { cwd?: string | URL | undefined; stdio?: unknown; encoding?: unknown; timeout?: number | undefined; env?: NodeJS.ProcessEnv | undefined },
   ) => string,
+  gateAuditFsLayer?: PublishGateAuditFsLayer,
+  /**
+   * AM-5 (feature publish-gate-audit-durable): test seam for the sibling-drift gate's `npm pack
+   * --dry-run --json` call — production leaves it unset (real `execFileSync`). Takes the package
+   * dir, returns raw stdout, or THROWS to simulate a real `npm` failure — a test can then prove the
+   * failure reaches `parseNpmPackInventory`'s caller as `unavailable`, never a real subprocess.
+   */
+  npmPackRunner?: (dir: string) => string,
 ): number {
   const json = flags.has('json');
   // Under --json stdout carries exactly one JSON document, so every human line — guard notes, refusals,
@@ -7187,19 +7344,55 @@ function cmdPublish(
       }
     });
 
-  // AM-6: an override (--allow-sibling-drift) is only real once its audit row is DURABLE. A write
-  // failure must refuse the publish rather than print "(logged)" about a log entry that never
-  // landed — the same "absence of a receipt is not success" lesson the registry-probe gate already
-  // enforces for a publish's own confirmation.
-  const auditedOverride = (detail: string, humanMessage: string, pkgNameForBlock: string): boolean => {
-    const wrote = appendPublishGateAudit(cwd, 'sibling-drift', 'warn', detail, '--allow-sibling-drift');
-    if (wrote) {
-      write(`dz publish: ⚠ ${humanMessage} — allowed via --allow-sibling-drift (logged)`);
-      return false;
+  // AM-4: `npm pack --dry-run --json` is a real subprocess — cache it for the lifetime of this
+  // ENTIRE run (keyed by resolved dir), NOT per package being checked (round-1 review, finding 5):
+  // the cache used to be re-created inside the per-package loop body, so two different dependents
+  // of the SAME sibling packed it twice. `npmPackRunner` (AM-5) is a test seam — production leaves
+  // it unset and runs the real subprocess; a test injects a stub that throws to prove a real `npm`
+  // failure reaches the caller as `unavailable`, without spawning anything.
+  // Lead fix after the fix-round's live dry-run (2026-09-14 01:02, MEASURED on the hub): the
+  // workspace side is now PACKED BY THE LIVE TRANSPORT — `pnpm pack` into a per-run temp dir,
+  // unpacked, and handed to core as a `packedDir` that core hashes with the SAME full walk it uses
+  // for the published tarball. `npm pack --dry-run --json` (kept behind the `npmPackRunner` test
+  // seam) never lists the LICENSE pnpm synthesises from the workspace root into a package whose own
+  // tree has none, so two siblings unchanged since publication (harness-presets, scout) read as
+  // "LICENSE only in the published copy" — a false drift the fix-round's inventory could not see.
+  // Honest limit: the seam path (tests) still parses npm's JSON; only production takes the pnpm path.
+  const npmPackInventoryCache = new Map<string, LocalInventoryResult>();
+  let packTmpDir: string | undefined;
+  const npmPackInventory = (dir: string): LocalInventoryResult => {
+    const key = resolve(dir);
+    const hit = npmPackInventoryCache.get(key);
+    if (hit !== undefined) return hit;
+    let out: LocalInventoryResult;
+    try {
+      if (npmPackRunner !== undefined) {
+        out = parseNpmPackInventory(npmPackRunner(dir));
+      } else {
+        packTmpDir ??= mkdtempSync(join(tmpdir(), 'dz-drift-pack-'));
+        out = { packedDir: extractIntoTempDir(dir, mkdtempSync(join(packTmpDir, 'p-'))).dir };
+      }
+    } catch (err) {
+      const how = npmPackRunner !== undefined ? 'npm pack --dry-run --json' : 'pnpm pack';
+      out = { unavailable: `${how} failed: ${(err as Error).message.split('\n')[0]}` };
     }
-    write(`dz publish: BLOCKED ${pkgNameForBlock} — ${humanMessage}, and the override could not be recorded (audit write failed); refusing rather than proceeding unlogged`);
-    return true;
+    npmPackInventoryCache.set(key, out);
+    return out;
   };
+  const localInventorySource: InventorySource = npmPackRunner !== undefined ? 'npm-pack' : 'pnpm-pack';
+
+  // AM-3 (Codex round-1 review, finding 4, high): sibling-drift audit records are EXACTLY one per
+  // package per rule per RUN. The old code appended one JSONL record per SIBLING a package depends
+  // on (a package with two drifted deps wrote two rows under the same rule), and the `unavailable`
+  // branch without `--allow-sibling-drift` wrote NO record at all. Every sibling outcome for a
+  // package is now aggregated first (`pkParts`/`pkVerdict`/`pkOverrideUsed`) and written ONCE —
+  // `block` if any sibling blocks, else `warn` if the only issues were resolved via
+  // `--allow-sibling-drift`, else `pass` — with a detail naming every sibling and its status.
+  // `siblingDriftAudited` guarantees the single write even though `--include-drifted`'s fixed-point
+  // loop can revisit the SAME package across rounds: a package's own `dependencies` never change
+  // between rounds, so a later round can only ever re-derive a SUBSET of what the first pass
+  // already covered (its siblings that drifted got folded into the batch and are now skipped).
+  const siblingDriftAudited = new Set<string>();
 
   let driftBlocked = 0;
   const extraBatch = new Set<string>();
@@ -7208,6 +7401,9 @@ function cmdPublish(
   // (finding 2) showed the single pass never re-checked an EXPANDED batch's own new edges. Capped at
   // `allPackages.length + 1` rounds (the plan's own "цикл с потолком = число пакетов").
   const maxRounds = allPackages.length + 1;
+  // Codex round-3 (2026-09-14): the per-run pack scratch is released in a `finally`, so an
+  // exception thrown while hashing or auditing cannot leak a `dz-drift-pack-*` dir under tmpdir.
+  try {
   for (let round = 0; round < maxRounds; round++) {
     let addedThisRound = false;
     for (const pk of targets) {
@@ -7221,13 +7417,25 @@ function cmdPublish(
       } catch (err) {
         // AM-3: an unreadable/invalid package.json for a BATCH package is an input this HARD gate
         // cannot build — it must BLOCK, never silently degrade to "no dependencies" (which used to
-        // read as a clean n/a).
+        // read as a clean n/a). AND (finding 4) the non-override branch below used to print a
+        // BLOCKED line with NO audit record behind it — an `unavailable` outcome is logged exactly
+        // like every other outcome, override or not.
         const reason = `package.json unreadable/invalid (${(err as Error).message.split('\n')[0]})`;
-        if (allowSiblingDrift) {
-          if (auditedOverride(`${pk.name}: ${reason}`, `sibling drift check unavailable for ${pk.name} (${reason})`, pk.name)) driftBlocked++;
-        } else {
-          write(`dz publish: BLOCKED ${pk.name} — sibling drift check unavailable (${reason}); add --allow-sibling-drift to override (logged) or fix the manifest`);
-          driftBlocked++;
+        if (!siblingDriftAudited.has(pk.name)) {
+          siblingDriftAudited.add(pk.name);
+          if (allowSiblingDrift) {
+            const wrote = appendPublishGateAudit(cwd, 'sibling-drift', 'warn', `${pk.name}: ${reason} — allowed via --allow-sibling-drift`, [{ name: pk.name, version: pk.version }], '--allow-sibling-drift', gateAuditFsLayer);
+            if (wrote.logged) {
+              write(`dz publish: ⚠ sibling drift check unavailable for ${pk.name} (${reason}) — allowed via --allow-sibling-drift (logged)`);
+            } else {
+              write(`dz publish: BLOCKED ${pk.name} — sibling drift check unavailable for ${pk.name} (${reason}), and the override could not be recorded (audit write failed: ${wrote.reason ?? 'unknown reason'}); refusing rather than proceeding unlogged`);
+              driftBlocked++;
+            }
+          } else {
+            const wrote = appendPublishGateAudit(cwd, 'sibling-drift', 'block', `${pk.name}: ${reason}`, [{ name: pk.name, version: pk.version }], undefined, gateAuditFsLayer);
+            write(`dz publish: BLOCKED ${pk.name} — sibling drift check unavailable (${reason}); add --allow-sibling-drift to override (logged) or fix the manifest${auditSuffix(wrote)}`);
+            driftBlocked++;
+          }
         }
         continue;
       }
@@ -7235,49 +7443,92 @@ function cmdPublish(
       const peerDeps = manifestObj?.peerDependencies ?? {};
       const optionalDeps = manifestObj?.optionalDependencies ?? {};
 
-      // AM-6: a package with no workspace: dependency at all is n/a for THIS gate — recorded as a
-      // pass note, not silence (FR-6 compatibility: output stays unchanged for such a batch).
+      // AM-6: a package with no EXTERNAL sibling to check at all is n/a for THIS gate — recorded as
+      // a pass note ("no external siblings"), not silence (FR-6 compatibility: this branch prints
+      // nothing to stdout, matching the pre-existing behavior). "External" covers BOTH "no
+      // workspace: dependency declared" and "every workspace: dependency is inside THIS batch"
+      // (publishing fresh, nothing stale to drift from) — both used to leave this package with no
+      // audit record at all when every dep resolved to the second case.
       const anyWorkspaceDep = [...Object.values(deps), ...Object.values(peerDeps), ...Object.values(optionalDeps)]
         .some((spec) => String(spec).startsWith('workspace:'));
-      if (!anyWorkspaceDep) {
-        appendPublishGateAudit(cwd, 'sibling-drift', 'pass', `${pk.name}: n/a — no workspace: dependency declared`);
-        continue;
-      }
 
-      const drifts = detectSiblingDrift({
-        dependencies: deps,
-        peerDependencies: peerDeps,
-        optionalDependencies: optionalDeps,
-        workspaceVersions,
-        workspaceDirs,
-        batch: batchNames,
-        fetchPublished,
-      });
+      const pkParts: string[] = [];
+      let pkVerdict: 'pass' | 'warn' | 'block' = 'pass';
+      let pkOverrideUsed = false;
+
+      if (anyWorkspaceDep) {
+        const drifts = detectSiblingDrift({
+          localInventory: npmPackInventory,
+          localInventorySource,
+          dependencies: deps,
+          peerDependencies: peerDeps,
+          optionalDependencies: optionalDeps,
+          workspaceVersions,
+          workspaceDirs,
+          batch: batchNames,
+          fetchPublished,
+        });
 
-      for (const r of drifts) {
-        if (r.status === 'same') {
-          appendPublishGateAudit(cwd, 'sibling-drift', 'pass', `${r.name}@${r.version} = workspace (dependent: ${pk.name})`);
-          write(`dz publish: ✓ sibling drift: none (${r.name}@${r.version} = workspace)`);
-        } else if (r.status === 'unavailable') {
-          if (allowSiblingDrift) {
-            if (auditedOverride(`${r.name}@${r.version}: ${r.reason}`, `sibling drift check unavailable for ${r.name}@${r.version} (${r.reason})`, pk.name)) driftBlocked++;
+        for (const r of drifts) {
+          if (r.status === 'same') {
+            pkParts.push(`${r.name}@${r.version}: same`);
+            write(`dz publish: ✓ sibling drift: none (${r.name}@${r.version} = workspace)`);
+          } else if (r.status === 'unavailable') {
+            if (allowSiblingDrift) {
+              pkParts.push(`${r.name}@${r.version}: unavailable (${r.reason}) — allowed via --allow-sibling-drift`);
+              if (pkVerdict !== 'block') pkVerdict = 'warn';
+              pkOverrideUsed = true;
+            } else {
+              pkParts.push(`${r.name}@${r.version}: unavailable (${r.reason})`);
+              pkVerdict = 'block';
+              write(`dz publish: BLOCKED ${pk.name} — sibling drift check unavailable (${r.reason}); add --allow-sibling-drift to override (logged) or check network/registry access`);
+              driftBlocked++;
+            }
+          } else if (includeDrifted) {
+            pkParts.push(`${r.name}@${r.version}: drift (${r.changedFiles.length} file(s)) — auto-included via --include-drifted`);
+            if (!batchNames.has(r.name) && !extraBatch.has(r.name)) {
+              extraBatch.add(r.name);
+              addedThisRound = true;
+              write(`dz publish: → sibling drift: ${r.name}@${r.version} differs from the workspace (${r.changedFiles.length} file(s)) — adding to the batch via --include-drifted${r.missingExports.length > 0 ? ` (missing exports: ${r.missingExports.join(', ')})` : ''}`);
+            }
+          } else if (allowSiblingDrift) {
+            pkParts.push(`${r.name}@${r.version}: drift (${r.changedFiles.length} file(s)) — allowed via --allow-sibling-drift`);
+            if (pkVerdict !== 'block') pkVerdict = 'warn';
+            pkOverrideUsed = true;
           } else {
-            write(`dz publish: BLOCKED ${pk.name} — sibling drift check unavailable (${r.reason}); add --allow-sibling-drift to override (logged) or check network/registry access`);
+            pkParts.push(`${r.name}@${r.version}: drift (${r.changedFiles.length} file(s))`);
+            pkVerdict = 'block';
+            const suggestFilter = filterStr !== undefined ? `${filterStr},${r.name}` : `${pk.name},${r.name}`;
+            write(`dz publish: BLOCKED ${pk.name} — sibling drift: @dzhechkov/${r.name.replace(/^@dzhechkov\//, '')}@${r.version} on the registry differs from the workspace (${r.changedFiles.length} file(s)); add ${r.name} to the batch (--filter ${suggestFilter}) or publish it first`);
             driftBlocked++;
           }
-        } else if (includeDrifted) {
-          if (!batchNames.has(r.name) && !extraBatch.has(r.name)) {
-            extraBatch.add(r.name);
-            addedThisRound = true;
-            write(`dz publish: → sibling drift: ${r.name}@${r.version} differs from the workspace (${r.changedFiles.length} file(s)) — adding to the batch via --include-drifted${r.missingExports.length > 0 ? ` (missing exports: ${r.missingExports.join(', ')})` : ''}`);
+        }
+      }
+
+      // AM-3/AM-6: the single, aggregated audit write for this package — "no external siblings"
+      // when nothing was ever checked, otherwise every sibling's status joined into one detail.
+      if (!siblingDriftAudited.has(pk.name)) {
+        siblingDriftAudited.add(pk.name);
+        const detail = pkParts.length > 0 ? pkParts.join('; ') : 'no external siblings';
+        // AM-6: an override reason is attached only when the FINAL verdict is 'warn' — if some
+        // OTHER sibling still stands as a live block, the override never actually excused the run.
+        const overrideReason = pkOverrideUsed && pkVerdict !== 'block' ? '--allow-sibling-drift' : undefined;
+        const wrote = appendPublishGateAudit(cwd, 'sibling-drift', pkVerdict, detail, [{ name: pk.name, version: pk.version }], overrideReason, gateAuditFsLayer);
+        if (pkOverrideUsed) {
+          // AM-6: the override is only real once ITS audit row is durable — a write failure must
+          // refuse the publish rather than print "(logged)" about a record that never landed.
+          if (wrote.logged) {
+            write(`dz publish: ⚠ sibling drift override recorded for ${pk.name} — allowed via --allow-sibling-drift (logged)`);
+          } else {
+            write(`dz publish: BLOCKED ${pk.name} — the --allow-sibling-drift override could not be recorded (audit write failed: ${wrote.reason ?? 'unknown reason'}); refusing rather than proceeding unlogged`);
+            driftBlocked++;
           }
-        } else if (allowSiblingDrift) {
-          if (auditedOverride(`${r.name}@${r.version}: ${r.changedFiles.length} file(s) differ from the workspace`, `sibling drift: ${r.name}@${r.version} differs from the workspace (${r.changedFiles.length} file(s))`, pk.name)) driftBlocked++;
-        } else {
-          appendPublishGateAudit(cwd, 'sibling-drift', 'block', `${pk.name} depends on ${r.name}@${r.version}; ${r.changedFiles.length} file(s) differ from the workspace`);
-          const suggestFilter = filterStr !== undefined ? `${filterStr},${r.name}` : `${pk.name},${r.name}`;
-          write(`dz publish: BLOCKED ${pk.name} — sibling drift: @dzhechkov/${r.name.replace(/^@dzhechkov\//, '')}@${r.version} on the registry differs from the workspace (${r.changedFiles.length} file(s)); add ${r.name} to the batch (--filter ${suggestFilter}) or publish it first`);
-          driftBlocked++;
+        } else if (pkParts.length > 0) {
+          write(`dz publish: ℹ sibling drift audit for ${pk.name}${auditSuffix(wrote)}`);
+        } else if (!wrote.logged) {
+          // Codex round-2 (2026-09-14), AM-6 residual: the "no external siblings" pass note stays
+          // silent on stdout ONLY while its record actually landed — a failed audit write is said.
+          write(`dz publish: ℹ sibling drift audit for ${pk.name} (no external siblings)${auditSuffix(wrote)}`);
         }
       }
     }
@@ -7292,6 +7543,9 @@ function cmdPublish(
     targets = allPackages.filter(matchesFilter);
     batchNames = new Set(targets.map((p) => p.name));
   }
+  } finally {
+    if (packTmpDir !== undefined) rmSync(packTmpDir, { recursive: true, force: true });
+  }
 
   const siblingDriftFailed = driftBlocked > 0;
   if (siblingDriftFailed && !dryRun) {
@@ -7328,8 +7582,8 @@ function cmdPublish(
   let packedInstallSmokePreviewFailed = false;
   if (dryRun) {
     if (bins.length === 0) {
-      appendPublishGateAudit(cwd, 'packed-install-smoke', 'pass', 'n/a — nothing in the batch declares a bin');
-      write('dz publish: ○ packed install smoke: n/a (nothing in the batch declares a bin)');
+      const wrote = appendPublishGateAudit(cwd, 'packed-install-smoke', 'pass', 'n/a — nothing in the batch declares a bin', targets.map((p) => ({ name: p.name, version: p.version })), undefined, gateAuditFsLayer);
+      write(`dz publish: ○ packed install smoke: n/a (nothing in the batch declares a bin)${auditSuffix(wrote)}`);
     } else {
       const scratchRoot = packedInstallScratchRoot();
       const packDir = mkdtempSync(join(scratchRoot, 'dz-publish-pack-'));
@@ -7361,36 +7615,30 @@ function cmdPublish(
       // Lead edit after the live dry-run (13.09 12:05): the preview packed the WORKING directory with
       // `workspace:^` specs still inside, so `npm install <tgz>` died with EUNSUPPORTEDPROTOCOL — the
       // preview must stage package.json exactly as the live packedTransport does (sibling pins via
-      // rewriteWorkspaceSpecs, prepublishOnly dropped) and restore the originals afterwards.
-      const stagedOriginals: Array<{ path: string; text: string }> = [];
+      // rewriteWorkspaceSpecs, prepublishOnly dropped) and restore the originals afterwards. Shared
+      // with `dz release` via `withStagedPackageJson` (feature release-smoke-staged-pack).
       try {
-        for (const p of targets) {
-          const pkgJsonPath = join(p.dir, 'package.json');
-          const original = readFileSync(pkgJsonPath, 'utf-8');
-          const rewritten = JSON.parse(rewriteWorkspaceSpecs(original, workspaceVersions)) as Record<string, unknown>;
-          const scripts = rewritten['scripts'];
-          if (scripts !== null && typeof scripts === 'object' && !Array.isArray(scripts)) delete (scripts as Record<string, unknown>)['prepublishOnly'];
-          stagedOriginals.push({ path: pkgJsonPath, text: original });
-          writeFileSync(pkgJsonPath, JSON.stringify(rewritten, null, 2) + '\n');
-        }
-        for (const step of smokePlan.steps) {
-          const r = runSmoke(step.cmd, { cwd: step.cwd, timeoutMs: step.timeoutMs });
-          smokeExecutions.push({ stepId: step.id, exitCode: r.exitCode, stdout: r.stdout, stderr: r.stderr, ...(r.timedOut !== undefined ? { timedOut: r.timedOut } : {}) });
-        }
-      } finally {
-        for (const o of stagedOriginals) { try { writeFileSync(o.path, o.text); } catch (err) { write(`dz publish: ⚠ could not restore ${o.path} after the preview smoke: ${formatPublishError(err)}`); } }
+        withStagedPackageJson(targets, workspaceVersions, write, () => {
+          for (const step of smokePlan.steps) {
+            const r = runSmoke(step.cmd, { cwd: step.cwd, timeoutMs: step.timeoutMs });
+            smokeExecutions.push({ stepId: step.id, exitCode: r.exitCode, stdout: r.stdout, stderr: r.stderr, ...(r.timedOut !== undefined ? { timedOut: r.timedOut } : {}) });
+          }
+        }, 'dz publish');
+      } catch (err) {
+        // a restore failure is a preview failure (Codex finding 1): never a green preview on a staged tree
+        smokeExecutions.push({ stepId: 'staged-restore', exitCode: 1, stdout: '', stderr: formatPublishError(err) });
       }
       const smokeVerdict = judgePackedInstallSmoke(smokePlan, smokeExecutions);
       try { rmSync(packDir, { recursive: true, force: true }); } catch { /* best-effort cleanup */ }
       try { rmSync(installDir, { recursive: true, force: true }); } catch { /* best-effort cleanup */ }
 
       if (smokeVerdict.ok) {
-        appendPublishGateAudit(cwd, 'packed-install-smoke', 'pass', 'preview: pack/install/--version all clean');
-        write('dz publish: ✓ packed install smoke (preview)');
+        const wrote = appendPublishGateAudit(cwd, 'packed-install-smoke', 'pass', 'preview: pack/install/--version all clean', targets.map((p) => ({ name: p.name, version: p.version })), undefined, gateAuditFsLayer);
+        write(`dz publish: ✓ packed install smoke (preview)${auditSuffix(wrote)}`);
       } else {
         const detail = smokeVerdict.failureDetail ?? smokeVerdict.bins.find((b) => !b.ok)?.detail ?? '(no detail)';
-        appendPublishGateAudit(cwd, 'packed-install-smoke', 'block', detail);
-        write(`dz publish: BLOCKED — packed install smoke failed (preview): ${detail}`);
+        const wrote = appendPublishGateAudit(cwd, 'packed-install-smoke', 'block', detail, targets.map((p) => ({ name: p.name, version: p.version })), undefined, gateAuditFsLayer);
+        write(`dz publish: BLOCKED — packed install smoke failed (preview): ${detail}${auditSuffix(wrote)}`);
         for (const b of smokeVerdict.bins.filter((b) => !b.ok)) write(`  ✗ ${b.pkg} (${b.binName}): ${b.detail ?? '(no detail)'}`);
         packedInstallSmokePreviewFailed = true;
       }
@@ -7453,22 +7701,35 @@ function cmdPublish(
         // `pnpm publish` and skipped re-signing (ADR-001, features/publish-gate-verifies-the-tarball).
         let cleanupGate: (() => void) | null = null;
         try {
-          const signed = JSON.parse(readFileSync(manifestPath, 'utf8'));
-          let extracted: { dir: string; cleanup: () => void } | undefined;
-          try {
-            extracted = extractPublishTarball(pk.dir);
-            cleanupGate = extracted.cleanup;
-          } catch (err) {
-            // Cross-family review (codex `gpt-5.6-sol`, 2026-08-22): falling back to the working
-            // TREE here fails the gate OPEN. The gate's whole claim is "what ships matches the
-            // signature"; with no artifact, nothing was compared, and reporting a pass would be a
-            // claim about an object that was never built. Say why, and block.
-            write(`dz publish: could not pack ${pk.name} (${(err as Error).message.split('\n')[0]}) — the artifact was never built, so its signature was not checked`);
+          const parsedManifest = JSON.parse(readFileSync(manifestPath, 'utf8'));
+          // FR-4 (feature publish-gate-audit-durable): a manifest that PARSES but is not a plain
+          // object — `null`, an array, a bare string — is UNAVAILABLE, never "no signature". Those
+          // are different failures with different fixes: "no signature" means run `dz sign`; a
+          // malformed manifest means the file itself is corrupt/wrong-shaped and re-signing alone
+          // would silently paper over that. Mirrors `readManifest`'s shape guard in
+          // publish-sibling-drift.ts (null/array/non-object ⇒ cannot be used, say so).
+          if (parsedManifest === null || typeof parsedManifest !== 'object' || Array.isArray(parsedManifest)) {
+            const gotShape = parsedManifest === null ? 'null' : Array.isArray(parsedManifest) ? 'an array' : typeof parsedManifest;
+            write(`dz publish: ${pk.name}'s ${MANIFEST_NAME} is not a JSON object (got ${gotShape}) — its signature is unavailable, not merely absent`);
             artifactUnavailable = true;
+          } else {
+            const signed = parsedManifest;
+            let extracted: { dir: string; cleanup: () => void } | undefined;
+            try {
+              extracted = extractPublishTarball(pk.dir);
+              cleanupGate = extracted.cleanup;
+            } catch (err) {
+              // Cross-family review (codex `gpt-5.6-sol`, 2026-08-22): falling back to the working
+              // TREE here fails the gate OPEN. The gate's whole claim is "what ships matches the
+              // signature"; with no artifact, nothing was compared, and reporting a pass would be a
+              // claim about an object that was never built. Say why, and block.
+              write(`dz publish: could not pack ${pk.name} (${(err as Error).message.split('\n')[0]}) — the artifact was never built, so its signature was not checked`);
+              artifactUnavailable = true;
+            }
+            verifyOk =
+              extracted !== undefined &&
+              verifyManifest(extracted.dir, signed, readFileSync(trustRoot, 'utf8')).ok;
           }
-          verifyOk =
-            extracted !== undefined &&
-            verifyManifest(extracted.dir, signed, readFileSync(trustRoot, 'utf8')).ok;
         } catch {
           verifyOk = false;
         } finally {
@@ -7513,9 +7774,10 @@ function cmdPublish(
       // batch resolves to.
       smoke: (artifacts): { ok: boolean; reason?: string } => {
         for (const a of artifacts) write(`dz publish: tarball ${a.name}@${a.newVersion} sha256:${a.sha256}`);
+        const auditPackages = artifacts.map((a) => ({ name: a.name, version: a.newVersion, tarballSha256: a.sha256 }));
         if (bins.length === 0) {
-          appendPublishGateAudit(cwd, 'packed-install-smoke', 'pass', 'n/a — nothing in the batch declares a bin');
-      write('dz publish: ○ packed install smoke: n/a (nothing in the batch declares a bin)');
+          const wrote = appendPublishGateAudit(cwd, 'packed-install-smoke', 'pass', 'n/a — nothing in the batch declares a bin', auditPackages, undefined, gateAuditFsLayer);
+          write(`dz publish: ○ packed install smoke: n/a (nothing in the batch declares a bin)${auditSuffix(wrote)}`);
           return { ok: true };
         }
         const scratchRoot = packedInstallScratchRoot();
@@ -7554,13 +7816,13 @@ function cmdPublish(
         const verdict = judgePackedInstallSmoke(smokePlan, smokeExecutions);
         try { rmSync(installDir, { recursive: true, force: true }); } catch { /* best-effort cleanup */ }
         if (verdict.ok) {
-          appendPublishGateAudit(cwd, 'packed-install-smoke', 'pass', 'pack/install/--version all clean (live, packedTransport)');
-          write('dz publish: ✓ packed install smoke');
+          const wrote = appendPublishGateAudit(cwd, 'packed-install-smoke', 'pass', 'pack/install/--version all clean (live, packedTransport)', auditPackages, undefined, gateAuditFsLayer);
+          write(`dz publish: ✓ packed install smoke${auditSuffix(wrote)}`);
           return { ok: true };
         }
         const detail = verdict.failureDetail ?? verdict.bins.find((b) => !b.ok)?.detail ?? '(no detail)';
-        appendPublishGateAudit(cwd, 'packed-install-smoke', 'block', detail);
-        write(`dz publish: BLOCKED — packed install smoke failed: ${detail}`);
+        const wrote = appendPublishGateAudit(cwd, 'packed-install-smoke', 'block', detail, auditPackages, undefined, gateAuditFsLayer);
+        write(`dz publish: BLOCKED — packed install smoke failed: ${detail}${auditSuffix(wrote)}`);
         for (const b of verdict.bins.filter((b) => !b.ok)) write(`  ✗ ${b.pkg} (${b.binName}): ${b.detail ?? '(no detail)'}`);
         return { ok: false, reason: detail };
       },
@@ -7790,7 +8052,7 @@ function cmdPublish(
 /*  ADR-001): computed from the declarative model, never hand-written  */
 /* ------------------------------------------------------------------ */
 
-function cmdParity(options: Map<string, string>, flags: Set<string>, write: Write, writeErr: WriteErr, cwd: string): number {
+async function cmdParity(options: Map<string, string>, flags: Set<string>, write: Write, writeErr: WriteErr, cwd: string): Promise<number> {
   const json = flags.has('json');
   if (flags.has('help')) {
     write('dz parity [--target <name>] [--json] — the computed feature×target map (never hand-written)');
@@ -7818,30 +8080,53 @@ function cmdParity(options: Map<string, string>, flags: Set<string>, write: Writ
     }
   }
 
-  // ADR-001 Decision 3 (feature setup-installs-apply-leg): `learning-apply` on `claude-code` is
-  // MEASURED, not declared — `hooks-prompt` is present for that ONE target only when
-  // `applyLegStatus(root).installed`. `computeParity` itself is untouched (FR-5); only the
-  // capability SET fed into it for this one cell differs from the static `TARGET_CAPABILITIES`.
-  // `applyLegStatus` never throws (fix round 1, Q3 finding: an unreadable helper used to be able to
-  // crash this command rather than degrade to a named remedy).
+  // ADR-001 Decision 3 (feature setup-installs-apply-leg), extended by `apply-leg-never-silent`
+  // Decision 3 (FR-4): `learning-apply` on `claude-code` is MEASURED, not declared — `hooks-prompt`
+  // is present for that ONE target only when the leg is OBSERVED to inject, not merely installed.
+  // Issue #2 was exactly this cell reading `full` while the leg injected nothing in every session
+  // but one: `applyLegStatus(root).installed` alone (file presence + structural wiring) is
+  // necessary but not sufficient — `probeApplyLeg` is the live end-to-end proof. `computeParity`
+  // itself is untouched (FR-5 of the earlier feature); only the capability SET fed into it for this
+  // one cell differs from the static `TARGET_CAPABILITIES`, exactly as before this feature.
+  // `applyLegStatus` never throws (fix round 1, Q3 finding); `probeApplyLeg` is only invoked when
+  // `installed` is true — a leg with missing/stale/unreadable helpers has nothing live to probe,
+  // and the pre-existing remedy for that state is unchanged.
   const applyLegStatusVal = applyLegStatus(cwd);
   const applyLegInstalled = applyLegStatusVal.installed;
+  // Codex round-2: `probeApplyLeg` may REJECT (temp dir, dynamic import, a throwing remover) — doctor
+  // catches that; parity must too, or a probe crash would crash `dz parity` instead of reading as
+  // "installed but silent: probe-error".
+  let applyLegProbe: Awaited<ReturnType<typeof probeApplyLeg>> | undefined;
+  if (applyLegInstalled) {
+    try {
+      applyLegProbe = await probeApplyLeg(cwd);
+    } catch (err) {
+      applyLegProbe = { ok: false, reason: `probe-error: ${err instanceof Error ? err.message : String(err)}`, elapsedMs: 0 };
+    }
+  }
+  const applyLegWorking = applyLegProbe?.ok === true;
   const matrix = buildParityMatrix().map((row) => {
-    if (row.feature.id !== 'learning-apply' || applyLegInstalled) return row;
+    if (row.feature.id !== 'learning-apply' || applyLegWorking) return row;
     const claudeCodeCaps = TARGET_CAPABILITIES['claude-code'].filter((c) => c !== 'hooks-prompt');
     return { feature: row.feature, cells: { ...row.cells, 'claude-code': computeParity(row.feature, claudeCodeCaps) } };
   });
-  // The "not installed" remedy — named ONLY for the one cell whose grant is a live measurement,
-  // never a blanket note for every `manual` cell (most targets are manual by DESIGN, not absence).
-  // `stale-version`/`unreadable` route through `applyLegReasonMessage` — the SAME text-producing
-  // function `dz doctor` uses for those two reasons (fix round 1, HIGH finding 2 / Q3 finding 7), so
-  // the two instruments cannot disagree about WHY a stale or broken install is not "full".
+  // The remedy — named ONLY for the one cell whose grant is a live measurement, never a blanket
+  // note for every `manual` cell (most targets are manual by DESIGN, not absence). Three distinct
+  // states, never conflated: NOT installed (stale-version/unreadable route through
+  // `applyLegReasonMessage`, the SAME text-producing function `dz doctor` uses for those two
+  // reasons, so the two instruments cannot disagree about WHY); installed but the live probe did
+  // NOT observe injection (FR-4's new "installed but silent" case, reason from `probeApplyLeg`
+  // itself — the SAME reason `dz doctor`'s live-probe row prints, so doctor and parity cannot
+  // disagree about a dead leg either); and the working case, where this function returns ''.
   const applyLegRemedy = (featureId: string, t: TargetName): string => {
-    if (featureId !== 'learning-apply' || t !== 'claude-code' || applyLegInstalled) return '';
-    if (applyLegStatusVal.reason === 'stale-version' || applyLegStatusVal.reason === 'unreadable') {
-      return ` — ${applyLegReasonMessage(applyLegStatusVal)}`;
+    if (featureId !== 'learning-apply' || t !== 'claude-code' || applyLegWorking) return '';
+    if (!applyLegInstalled) {
+      if (applyLegStatusVal.reason === 'stale-version' || applyLegStatusVal.reason === 'unreadable') {
+        return ` — ${applyLegReasonMessage(applyLegStatusVal)}`;
+      }
+      return ' — not installed — run dz setup --target claude-code --memory agentdb';
     }
-    return ' — not installed — run dz setup --target claude-code --memory agentdb';
+    return ` — installed but silent: ${applyLegProbe?.reason ?? 'unknown'}`;
   };
   // EVIDENCE staleness, folded into the report (fix round 2, R2-3). Derived from the records
   // themselves — no `codex --version`, no subprocess, so `dz parity` stays a deterministic function
@@ -8116,7 +8401,24 @@ function cmdRelease(options: Map<string, string>, flags: Set<string>, cwd: strin
   let smokeTmp: string | undefined;
   const execSteps: GateStep[] = plan.steps.filter((s) => s.kind !== 'synthetic-fail');
   say(`\ndz release — executing ${execSteps.length} gate step(s) across ${plan.packages.length} package(s)…`);
-  for (const step of execSteps) {
+
+  // FR-1/FR-2 (feature release-smoke-staged-pack): the packed-install smoke's `pack` steps must
+  // pack the SAME staged bytes `dz publish`'s preview does — sibling `workspace:*` deps rewritten
+  // to the exact sibling version, `prepublishOnly` dropped — or `npm install` on the resulting
+  // tarballs dies with EUNSUPPORTEDPROTOCOL (MEASURED 2026-09-13 16:05). `workspaceVersions` is
+  // built like publish's (from the FULL workspace, not just this release's filtered batch — an
+  // out-of-batch sibling still needs its real version). Only the `pack` sub-steps are staged; the
+  // `install`/`bin-exists`/`bin-version` steps already run against the packed tarballs and need no
+  // staging.
+  const allPackages = discoverPackages(cwd);
+  const workspaceVersions = new Map(allPackages.map((p) => [p.name, p.version]));
+  const packStepIds = new Set(
+    (plan.packedInstallPlan?.steps ?? [])
+      .filter((s) => s.kind === 'pack')
+      .map((s) => `smoke:packed-install:${s.id}`),
+  );
+
+  const runGateStep = (step: GateStep): void => {
     let stepCwd = step.cwd;
     if (step.tempCwd === true) {
       // AM-4: boot bins in a throwaway cwd so an installer-style bin cannot mutate the workspace.
@@ -8133,7 +8435,35 @@ function cmdRelease(options: Map<string, string>, flags: Set<string>, cwd: strin
       durationMs: Date.now() - started,
       timedOut: r.timedOut,
     });
+  };
+
+  // Plan/execution order is load-bearing (see the plan/execution skew guard test) — steps run in
+  // exactly the order `plan.steps` lists them, one loop, no reordering. Only the `pack` steps are
+  // wrapped in the staged window, individually, so where they fall in that order never changes.
+  let announcedStagedPack = false;
+  for (const step of execSteps) {
+    if (packStepIds.has(step.id)) {
+      if (!announcedStagedPack) {
+        say('dz release: smoke: packed tarballs staged like the live publish (workspace:* → exact sibling versions)');
+        announcedStagedPack = true;
+      }
+      // Codex finding 2: stage ONLY the package this pack step packs — sibling pins come from
+      // workspaceVersions, and `npm pack` reads the packed package's manifest alone.
+      const packed = factsList.filter((f) => f.name === step.pkg);
+      try {
+        withStagedPackageJson(packed, workspaceVersions, write, () => runGateStep(step), 'dz release');
+      } catch (err) {
+        // a staging/restore failure is a FAILED step with the reason, never a silent skip — and ONE
+        // record per step: a run already recorded by runGateStep is replaced, not duplicated (Codex r2)
+        const failed = { stepId: step.id, exitCode: 1, stdout: '', stderr: formatPublishError(err), durationMs: 0, timedOut: false };
+        const at = executions.findIndex((e) => e.stepId === step.id);
+        if (at >= 0) executions[at] = failed; else executions.push(failed);
+      }
+    } else {
+      runGateStep(step);
+    }
   }
+
   if (smokeTmp !== undefined) {
     try { rmSync(smokeTmp, { recursive: true, force: true }); } catch { /* best-effort cleanup */ }
   }
@@ -8146,7 +8476,22 @@ function cmdRelease(options: Map<string, string>, flags: Set<string>, cwd: strin
   for (const g of verdict.gates) {
     const icon = g.status === 'pass' ? '✓' : g.status === 'fail' ? '✗' : '○';
     say(`  ${icon} ${g.gate.padEnd(7)} ${g.status.toUpperCase().padEnd(4)}  ${g.passed} passed, ${g.failures.length} failed, ${g.skips.length} skipped`);
-    for (const f of g.failures) say(`      [${f.class}] ${f.pkg !== undefined ? `${f.pkg}: ` : ''}${f.reason}`);
+    for (const f of g.failures) {
+      say(`      [${f.class}] ${f.pkg !== undefined ? `${f.pkg}: ` : ''}${f.reason}`);
+      // FR-2 / AM-4 / AM-6 (feature release-gate-output-tail): print each non-empty stream's
+      // tail under the failure line, labelled `stdout:`/`stderr:` at the failure's own 6-space
+      // indent, with its content lines at an 8-space CONTINUATION indent so a reader can tell a
+      // tail line from a new failure/skip bullet at a glance; --json carries the same `tails`
+      // object as-is (present, possibly with empty strings, on every executed failure).
+      if (f.tails !== undefined) {
+        for (const stream of ['stdout', 'stderr'] as const) {
+          const t = f.tails[stream];
+          if (t.length === 0) continue;
+          say(`      ${stream}:`);
+          for (const tailLine of t.split('\n')) say(`        ${tailLine}`);
+        }
+      }
+    }
     for (const sk of g.skips) say(`      [${sk.class}] ${sk.pkg}: ${sk.reason}`);
   }
   say(`  verdict at ${verdict.timestamp}: ${verdict.publishAction === 'proceed' ? '✓ all gates green' : '✗ RELEASE BLOCKED'}`);
@@ -10865,41 +11210,175 @@ function runGuardEvaluation(root: string, op: string, text: string | undefined,
   return result;
 }
 
+/** A package this audit call concerns (feature `publish-gate-audit-durable`, FR-1). */
+interface PublishGateAuditPackage {
+  readonly name: string;
+  readonly version: string;
+  /** Present once a tarball has actually been built (packedTransport/smoke); absent → `sha256:n/a`. */
+  readonly tarballSha256?: string;
+}
+
+/**
+ * fs primitives `appendPublishGateAudit` needs for its durable write (FR-2), injectable so a test
+ * can make `fsyncSync` throw without touching the real filesystem underneath every OTHER seam this
+ * function shares with production. Left unset in production → the real `node:fs` functions above.
+ */
+interface PublishGateAuditFsLayer {
+  readonly existsSync: (path: string) => boolean;
+  readonly mkdirSync: (path: string, opts: { recursive: boolean }) => void;
+  readonly openSync: (path: string, flags: number) => number;
+  /**
+   * AM-2 (Codex round-1 review, finding 2, high): `Buffer`, not `string` — a SHORT write must
+   * resume at the exact BYTE it stopped at, and a string-based API cannot express that safely once
+   * the data contains any multi-byte UTF-8 character (re-encoding a slice of an already-partial
+   * string can silently produce different bytes than the ones actually pending). The real
+   * `node:fs.writeSync` accepts a `Buffer` directly (no re-encoding), so this changes nothing about
+   * what production writes.
+   */
+  readonly writeSync: (fd: number, data: Buffer) => number;
+  readonly fsyncSync: (fd: number) => void;
+  readonly closeSync: (fd: number) => void;
+}
+
+/**
+ * AM-2: loop `writeSync` until every byte of `data` has been accepted, checking the RETURNED
+ * length on every call (Codex round-1 review, finding 1: the old code called `writeSync` once and
+ * ignored the return value — a short write followed by `fsyncSync` durably persists a TRUNCATED,
+ * unparseable JSON line into an append-only log every future read walks). A call that reports zero
+ * or negative progress can never complete the buffer and would spin forever — that is treated as a
+ * failure, not a retry target.
+ */
+function writeAllSync(fsLayer: PublishGateAuditFsLayer, fd: number, data: Buffer): void {
+  let remaining = data;
+  while (remaining.length > 0) {
+    const n = fsLayer.writeSync(fd, remaining);
+    if (typeof n !== 'number' || !Number.isFinite(n) || n <= 0) {
+      throw new Error(`writeSync made no progress (returned ${n}) with ${remaining.length} byte(s) still pending`);
+    }
+    // Codex round-2 (2026-09-14) finding 3: a layer claiming MORE bytes than it was given is lying
+    // about the record, and must not reach "logged: true" through a subarray that just goes empty.
+    if (n > remaining.length) {
+      throw new Error(`writeSync claimed ${n} byte(s) written but only ${remaining.length} were supplied`);
+    }
+    remaining = remaining.subarray(n);
+  }
+}
+
+const REAL_PUBLISH_GATE_AUDIT_FS: PublishGateAuditFsLayer = { existsSync, mkdirSync, openSync, writeSync, fsyncSync, closeSync };
+
 /**
- * Feature `publish-sibling-drift-gate` (FR-5/AM-6): both the sibling-drift and packed-install-smoke
+ * Feature `publish-sibling-drift-gate` (FR-5/AM-6), durability + per-package identity added by
+ * feature `publish-gate-audit-durable` (FR-1/FR-2): both the sibling-drift and packed-install-smoke
  * gates write to the SAME append-only, hash-chained `.dz/guard-audit.jsonl` the declarative
  * `dz guard` rules use — visibility for `dz guard promote`/`dz compounding` never depends on
  * which mechanism produced the finding. `pass` records go through as an informational `note`
  * (never a violation, so they can never flip the row's own verdict) so a clean check is ALSO on
  * the record, not just a block or an override (AM-6: "аудит без записи = не аудит").
  *
- * Returns whether the write actually landed. Most callers are best-effort (a write failure never
- * blocks a verdict already decided) — the one exception is an `--allow-sibling-drift` OVERRIDE,
- * whose caller MUST check this return value: an override is not real without a durable row behind
- * it (AM-6's load-bearing property — see `auditedOverride` in `cmdPublish`).
+ * FR-1: one JSONL record PER PACKAGE in `packages`, each naming the package, its version, and the
+ * tarball's sha256 (or the explicit `sha256:n/a` before a tarball exists — a dry-run preview). FR-2:
+ * the write is `openSync('a') → writeSync → fsyncSync(fd) → closeSync`, plus an `fsyncSync` of the
+ * `.dz` directory itself the one time this call CREATES it (a file's own fsync durably persists its
+ * bytes; the directory entry that makes the file findable after a crash needs its own fsync — the
+ * same lesson `integration-apply.ts`'s `fsyncDirectory` already encodes).
+ *
+ * Returns `{ logged, reason? }` — never a bare boolean, so a caller can print WHY a write failed,
+ * not just that it did. Most callers are best-effort (a write failure never blocks a verdict already
+ * decided) — the one exception is an `--allow-sibling-drift` OVERRIDE, whose caller MUST check
+ * `logged`: an override is not real without a durable row behind it (AM-6's load-bearing property —
+ * see `auditedOverride` in `cmdPublish`).
  */
 function appendPublishGateAudit(
   root: string,
   rule: 'sibling-drift' | 'packed-install-smoke',
   verdict: 'pass' | 'warn' | 'block',
   detail: string,
+  packages: readonly PublishGateAuditPackage[],
   overrideReason?: string,
-): boolean {
+  fsLayer: PublishGateAuditFsLayer = REAL_PUBLISH_GATE_AUDIT_FS,
+): { logged: boolean; reason?: string } {
+  if (packages.length === 0) return { logged: true }; // nothing to record about — not a failure
+  const dzDir = join(root, '.dz');
+  const auditPath = join(dzDir, 'guard-audit.jsonl');
+  const dzDirExisted = fsLayer.existsSync(dzDir);
+  // AM-2: tracked BEFORE the write — this is what decides whether the append call below is about
+  // to CREATE the file (needing a `.dz` directory-entry fsync afterwards) or extend an existing one
+  // (whose directory entry is already durable from a previous append).
+  const auditFileExisted = fsLayer.existsSync(auditPath);
   try {
-    const rec = auditRecord(
+    fsLayer.mkdirSync(dzDir, { recursive: true });
+  } catch (err) {
+    return { logged: false, reason: `could not create ${dzDir}: ${(err as Error).message.split('\n')[0]}` };
+  }
+  const records = packages.map((pkg) => {
+    const sha = pkg.tarballSha256 !== undefined && pkg.tarballSha256 !== '' ? pkg.tarballSha256 : 'n/a';
+    const label = `${rule}: ${pkg.name}@${pkg.version} sha256:${sha} — ${detail}`;
+    return auditRecord(
       verdict === 'pass'
-        ? { op: 'publish', verdict, violations: [], checked: [rule], notEstablished: [], notes: [`${rule}: ${detail}`] }
-        : { op: 'publish', verdict, violations: [{ rule, severity: 'hard', detail }], checked: [rule], notEstablished: [] },
+        ? { op: 'publish', verdict, violations: [], checked: [rule], notEstablished: [], notes: [label] }
+        : { op: 'publish', verdict, violations: [{ rule, severity: 'hard', detail: label }], checked: [rule], notEstablished: [] },
       new Date().toISOString(),
       overrideReason !== undefined ? { reason: overrideReason } : undefined,
     );
-    mkdirSync(join(root, '.dz'), { recursive: true });
-    const auditPath = join(root, '.dz', 'guard-audit.jsonl');
-    writeFileSync(auditPath, appendChainedLines([rec], readLogTail(auditPath)), { flag: 'a' });
-    return true;
-  } catch {
-    return false; // audit write failed — the caller decides whether that itself is refusable (AM-6)
+  });
+  let bytes: string;
+  try {
+    bytes = appendChainedLines(records, readLogTail(auditPath));
+  } catch (err) {
+    return { logged: false, reason: `could not build the chained record: ${(err as Error).message.split('\n')[0]}` };
   }
+  if (bytes === '') return { logged: true };
+  const buf = Buffer.from(bytes, 'utf-8');
+  let fd: number | undefined;
+  try {
+    fd = fsLayer.openSync(auditPath, fsConstants.O_APPEND | fsConstants.O_CREAT | fsConstants.O_WRONLY);
+    writeAllSync(fsLayer, fd, buf); // AM-2: loops on a short write, throws on no progress
+    fsLayer.fsyncSync(fd);
+    fsLayer.closeSync(fd);
+    fd = undefined;
+  } catch (err) {
+    return { logged: false, reason: (err as Error).message.split('\n')[0] ?? String(err) };
+  } finally {
+    if (fd !== undefined) { try { fsLayer.closeSync(fd); } catch { /* best-effort */ } }
+  }
+  // AM-2 (Codex round-1 review, finding 1, high): a directory-entry fsync is REQUIRED, not
+  // best-effort, whenever THIS call is the reason the entry needed persisting — swallowing its
+  // failure used to let `logged: true` go out about a record whose directory entry can vanish on a
+  // crash before the next fsck. Two DIFFERENT entries can need persisting, tracked independently:
+  if (!auditFileExisted) {
+    // The audit FILE was just created by the open() above — `.dz` (its containing directory) needs
+    // its own fsync so the new directory entry survives a crash; the file's own fsync (above) only
+    // guarantees the file's DATA, not that anything can find it afterwards.
+    let dfd: number | undefined;
+    try {
+      dfd = fsLayer.openSync(dzDir, fsConstants.O_RDONLY);
+      fsLayer.fsyncSync(dfd);
+    } catch (err) {
+      return { logged: false, reason: `could not fsync ${dzDir} after creating ${auditPath}: ${(err as Error).message.split('\n')[0]}` };
+    } finally {
+      if (dfd !== undefined) { try { fsLayer.closeSync(dfd); } catch { /* best-effort, does not affect the verdict already returned */ } }
+    }
+  }
+  if (!dzDirExisted) {
+    // `.dz` ITSELF was just created by `mkdirSync` above — its PARENT (`root`) needs its own fsync
+    // so `.dz`'s OWN directory entry survives a crash (the fsync of `.dz` just above only durably
+    // persists entries INSIDE `.dz`, not the fact that `.dz` exists at all).
+    let pfd: number | undefined;
+    try {
+      pfd = fsLayer.openSync(root, fsConstants.O_RDONLY);
+      fsLayer.fsyncSync(pfd);
+    } catch (err) {
+      return { logged: false, reason: `could not fsync ${root} after creating ${dzDir}: ${(err as Error).message.split('\n')[0]}` };
+    } finally {
+      if (pfd !== undefined) { try { fsLayer.closeSync(pfd); } catch { /* best-effort, does not affect the verdict already returned */ } }
+    }
+  }
+  return { logged: true };
+}
+
+/** FR-2: the printed suffix a caller appends to its own verdict line — never silent about logging. */
+function auditSuffix(wrote: { logged: boolean; reason?: string }): string {
+  return wrote.logged ? ' (logged)' : ` (audit NOT logged: ${wrote.reason ?? 'unknown reason'})`;
 }
 
 function renderGuardObservation(observation: GuardObservation): string {
@@ -12562,6 +13041,46 @@ function splitMutationGateOutputWrite(
   return 'path' in result ? { outputPath: result.path } : { outputError: result.error };
 }
 
+/** Fix-round 1 (AM-1, feature qe-step-gate-scoped-to-feature): a `--touched` path arrives in one of
+ * several shapes — package-relative POSIX (the common case, already correct as-is), with a leading
+ * `./`, as an ABSOLUTE path inside the package, or REPO-relative (e.g.
+ * `packages/@dzhechkov/harness-cli/src/x.ts`) — and on a POSIX host a caller may hand a backslash
+ * path too (a Windows-authored change list). MEASURED (Codex round-1 review): the pre-fix
+ * normalization only converted the native path separator, so any of the other shapes matched ZERO
+ * registry entries and the gate silently reported `selected 0` / exit 0 — the run looked clean while
+ * defending nothing. This resolves EVERY shape to the package-relative POSIX candidate(s) a registry
+ * entry's `file` is written in, and — when a path genuinely resolves outside `pkgDir` (an absolute
+ * path elsewhere, or a `../` that still escapes the package after a repo-relative reinterpretation)
+ * — reports it as `outside`, never a silent non-match indistinguishable from "the path doesn't
+ * exist".
+ */
+function normalizeTouchedPath(
+  raw: string,
+  pkgDir: string,
+  repoTop: string | null,
+): { readonly candidates: readonly string[]; readonly outside: boolean } {
+  // Codex round-2: normalize lexically FIRST so `lib/../../x` is seen as the traversal it is, and treat a
+  // Windows-drive path (`C:/…`) on a POSIX host as outside the package (counted, said) rather than as a
+  // package-relative name that silently matches nothing. Symlinked package dirs stay a named limit: the
+  // containment check is lexical, not realpath-based.
+  const p = posixNormalize(raw.replace(/\\/g, '/')).replace(/^(?:\.\/)+/, '');
+  if (p === '' || p === '.') return { candidates: [], outside: false };
+  if (/^[A-Za-z]:\//.test(p) && !isAbsolute(p)) return { candidates: [], outside: true };
+
+  if (isAbsolute(p)) {
+    const rel = relative(pkgDir, resolve(p)).split(sep).join('/');
+    return rel === '' || rel.startsWith('..') ? { candidates: [], outside: true } : { candidates: [rel], outside: false };
+  }
+
+  const candidates = new Set<string>();
+  if (p !== '..' && !p.startsWith('../')) candidates.add(p); // already package-relative, as given
+  if (repoTop !== null) {
+    const rel = relative(pkgDir, resolve(repoTop, p)).split(sep).join('/');
+    if (rel !== '' && !rel.startsWith('..')) candidates.add(rel);
+  }
+  return candidates.size > 0 ? { candidates: [...candidates], outside: false } : { candidates: [], outside: true };
+}
+
 function cmdMutationGate(
   options: Map<string, string>,
   flags: Set<string>,
@@ -12595,7 +13114,9 @@ function cmdMutationGate(
 
   let entries: readonly MutationRegistryEntry[] = parsed.registry.entries;
   let entryResults: readonly MutationEntryResult[] = parsed.entryResults;
+  const totalRegistryEntries = entries.length;
   const only = options.get('only');
+  let onlyIds: Set<string> | null = null;
   if (only !== undefined) {
     const ids = only.split(',').map((s) => s.trim()).filter(Boolean);
     const known = new Set([
@@ -12604,16 +13125,138 @@ function cmdMutationGate(
     ]);
     const unknown = ids.filter((id) => !known.has(id));
     if (unknown.length > 0) return fail(`--only names unknown entry id(s): ${unknown.join(', ')}`);
+    onlyIds = new Set(ids);
     entries = entries.filter((e) => ids.includes(e.id));
     entryResults = entryResults.filter((result) => ids.includes(result.id));
   }
 
+  // Feature qe-step-gate-scoped-to-feature (FR-1/FR-2/FR-3): scope the gate to the files a FEATURE
+  // actually touched and/or entries added since a base ref, instead of the whole registry — MEASURED
+  // 2026-09-12, 358 entries on this repo's core package ran 30-40 minutes and hit the timeout wall,
+  // INCONCLUSIVE every time, though a feature owns only its own touched files. `--touched` and
+  // `--added-since` UNION (a file changed by the feature OR an entry it newly added is in scope);
+  // that union then INTERSECTS with `--only` when both are given, same algebra as an ordinary filter
+  // chain. Both selectors are computed against the FULL (pre-`--only`) registry so their reported
+  // counts describe what THEY matched, independent of any `--only` narrowing applied on top.
+  const touchedRaw = options.get('touched');
+  const addedSinceRaw = options.get('added-since');
+  let selectionMeta: {
+    readonly selected: number;
+    readonly total: number;
+    readonly touched: readonly string[] | null;
+    readonly addedSince: readonly string[] | null;
+    readonly base: string | null;
+    readonly baseAbsent: boolean;
+    readonly outsidePackage: number;
+  } | null = null;
+  if (touchedRaw !== undefined || addedSinceRaw !== undefined) {
+    let repoTop: string | null = null;
+    try { repoTop = execSync('git rev-parse --show-toplevel', { cwd: pkgDir, stdio: 'pipe', encoding: 'utf-8' }).trim() || null; } catch { /* not in a git repo */ }
+
+    const touchedIds = new Set<string>();
+    let outsidePackageCount = 0;
+    if (touchedRaw !== undefined) {
+      const touchedPaths = touchedRaw.split(',').map((s) => s.trim()).filter(Boolean);
+      const normalizedCandidates = new Set<string>();
+      for (const raw of touchedPaths) {
+        const { candidates, outside } = normalizeTouchedPath(raw, pkgDir, repoTop);
+        if (outside) outsidePackageCount++;
+        for (const c of candidates) normalizedCandidates.add(c);
+      }
+      for (const entry of parsed.registry.entries) {
+        if (normalizedCandidates.has(entry.file)) touchedIds.add(entry.id);
+      }
+    }
+
+    const addedSinceIds = new Set<string>();
+    let baseAbsentMessage: string | null = null;
+    if (addedSinceRaw !== undefined) {
+      if (repoTop === null) return fail(`--added-since requires ${pkgDir} to be inside a git repository`);
+      try {
+        execFileSync('git', ['rev-parse', '--verify', '--quiet', `${addedSinceRaw}^{commit}`], { cwd: repoTop, stdio: 'pipe' });
+      } catch {
+        return fail(`--added-since names an unknown git ref: ${addedSinceRaw}`);
+      }
+      const registryRelPath = relative(repoTop, registryPath).split(sep).join('/');
+      let baseRegistry: MutationRegistry | null = null;
+      // Fix-round 1 (AM-3): only a genuinely ABSENT registry at that ref means "all entries are
+      // new". Any OTHER `git show` failure (a bad path that isn't a missing-registry case, git
+      // itself missing, a timeout) is a usage error, and a registry that parses to invalid JSON /
+      // an invalid registry shape at that ref is a distinct usage error too — collapsing all three
+      // into "absent" used to hide real failures behind a silently-too-generous selection.
+      try {
+        // stdio: registry-absent-at-ref is an EXPECTED outcome (git's own "fatal: path … does not
+        // exist" on stderr would otherwise leak to the terminal by node's default inherit-stderr
+        // behaviour) — pipe it into the caught error instead of printing it at the user.
+        const baseText = execFileSync('git', ['show', `${addedSinceRaw}:${registryRelPath}`], { cwd: repoTop, encoding: 'utf-8', env: { ...process.env, LC_ALL: 'C', LANG: 'C' }, maxBuffer: 16 * 1024 * 1024, stdio: ['ignore', 'pipe', 'pipe'] });
+        const baseParsed = parseMutationRegistry(baseText);
+        if (baseParsed.registry === null) {
+          return fail(`base registry at ${addedSinceRaw} is not a valid registry: ${baseParsed.errors.join('; ')}`);
+        }
+        baseRegistry = baseParsed.registry;
+      } catch (e) {
+        const stderr = typeof (e as { stderr?: unknown })?.stderr === 'string'
+          ? (e as { stderr: string }).stderr
+          : String((e as Error)?.message ?? e);
+        if (/does not exist in|exists on disk, but not in|Path .* does not exist/i.test(stderr)) {
+          baseRegistry = null; // registry absent at that ref → every current entry counts as added
+          baseAbsentMessage = `base registry absent at ${addedSinceRaw} — all ${totalRegistryEntries} entries count as added`;
+        } else {
+          return fail(`git show ${addedSinceRaw}:${registryRelPath} failed: ${stderr.trim()}`);
+        }
+      }
+      if (baseAbsentMessage !== null && !json) write(`mutation-gate: ${baseAbsentMessage}`);
+      for (const id of registryEntriesAddedSince(baseRegistry, parsed.registry)) addedSinceIds.add(id);
+    }
+
+    let unionIds = new Set<string>([...touchedIds, ...addedSinceIds]);
+    if (onlyIds !== null) {
+      const only2 = onlyIds;
+      unionIds = new Set([...unionIds].filter((id) => only2.has(id)));
+    }
+    const selectorParts: string[] = [];
+    if (touchedRaw !== undefined) selectorParts.push(`touched: ${touchedIds.size}${outsidePackageCount > 0 ? ` (${outsidePackageCount} outside package)` : ''}`);
+    if (addedSinceRaw !== undefined) selectorParts.push(`added-since ${addedSinceRaw}: ${addedSinceIds.size}`);
+    if (!json) write(`mutation-gate: selected ${unionIds.size} of ${totalRegistryEntries} entries (${selectorParts.join('; ')})`);
+    entries = entries.filter((e) => unionIds.has(e.id));
+    entryResults = entryResults.filter((result) => unionIds.has(result.id));
+
+    // AM-4: every scoped run — empty or not — carries a machine-readable `selection` object in the
+    // JSON contract, not just the empty-selection early return.
+    selectionMeta = {
+      selected: unionIds.size,
+      total: totalRegistryEntries,
+      touched: touchedRaw !== undefined ? [...touchedIds].sort() : null,
+      addedSince: addedSinceRaw !== undefined ? [...addedSinceIds].sort() : null,
+      base: addedSinceRaw ?? null,
+      baseAbsent: baseAbsentMessage !== null,
+      outsidePackage: outsidePackageCount,
+    };
+
+    if (unionIds.size === 0) {
+      const msg = `0 entries match ${selectorParts.join(', ')} — nothing to run`;
+      if (json) {
+        write(JSON.stringify({
+          packageDir: pkgDir,
+          registryPath,
+          selection: selectionMeta,
+          results: [],
+          summary: summarizeMutationResults([]),
+          exitCode: 0,
+        }, null, 2));
+      } else {
+        write(`mutation-gate: ${msg}`);
+      }
+      return 0;
+    }
+  }
+
   if (entries.length === 0) {
     const scope = only === undefined ? 'registry' : 'selected registry entries';
     const error = `${scope} has no runnable entries after validation — nothing can be run; the registry is unusable`;
     const summary = summarizeMutationResults(entryResults);
     if (json) {
-      write(JSON.stringify({ error, registryPath, results: entryResults, summary, exitCode: 2 }, null, 2));
+      write(JSON.stringify({ error, registryPath, ...(selectionMeta !== null ? { selection: selectionMeta } : {}), results: entryResults, summary, exitCode: 2 }, null, 2));
     } else {
       write(`dz mutation-gate: ${error}`);
       for (const result of entryResults) {
@@ -12625,7 +13268,7 @@ function cmdMutationGate(
 
   const testCmdRaw = options.get('test-cmd') ?? parsed.registry.testCommand ?? 'npm test';
   if (/[\0\n\r]/.test(testCmdRaw)) return fail('--test-cmd may not contain NUL or newline characters');
-  const testCmd = testCmdRaw;
+  let testCmd = testCmdRaw;
   const excludedSelfChecks = REGISTRY_SELFCHECK_TESTS.filter((testFile) =>
     entries.some((entry) => buildMutationTestCommand(testCmd, entry).excluded.includes(testFile)),
   );
@@ -12633,8 +13276,88 @@ function cmdMutationGate(
     write(`mutation-gate: self-check excluded from mutant runs: ${excludedSelfChecks.join(', ') || '(none)'}`);
   }
 
-  const timeoutOpt = Number(options.get('timeout') ?? '300000');
-  const timeout = Number.isFinite(timeoutOpt) && timeoutOpt > 0 ? timeoutOpt : 300000;
+  // mutation-gate-timeout-verdict FR-3/FR-4: precedence is the `--timeout` flag > the registry's
+  // own `timeoutMs` field > the 300000ms default. A package whose real baseline run is longer than
+  // the default (this repo's core package, MEASURED ≈5-8 min) declares its floor in the registry so
+  // a bare `dz mutation-gate` — no flag — still succeeds (AC-5).
+  const timeoutFlagRaw = options.get('timeout');
+  let timeout: number;
+  let timeoutSource: 'flag' | 'registry' | 'default';
+  if (timeoutFlagRaw !== undefined) {
+    const timeoutOpt = Number(timeoutFlagRaw);
+    if (Number.isFinite(timeoutOpt) && timeoutOpt > 0) {
+      timeout = timeoutOpt;
+      timeoutSource = 'flag';
+    } else {
+      timeout = 300000;
+      timeoutSource = 'default';
+    }
+  } else if (parsed.registry.timeoutMs !== undefined) {
+    timeout = parsed.registry.timeoutMs;
+    timeoutSource = 'registry';
+  } else {
+    timeout = 300000;
+    timeoutSource = 'default';
+  }
+  if (!json) write(`mutation-gate: timeout: ${timeout} ms (${timeoutSource})`);
+
+  // mutation-gate-baseline-honesty FR-2/FR-3: precedence is the `--max-workers` flag > the
+  // registry's own `maxWorkers` field > `min(4, max(1, floor(cpus/2)))`. Baseline and mutant runs
+  // spawn the package's FULL testCommand at vitest's DEFAULT worker count (= cpu cores) unless
+  // capped — MEASURED: under embedding-daemon tests (0.7-3.5 GB/process) this repo's core package
+  // hit load 62-358 and 0.4-1.8 GB free on an 8-core/16GB box, killing three full overnight runs
+  // (0bb74d66); the same suite at `--maxWorkers=2` passed 6909/6909.
+  const maxWorkersFlagRaw = options.get('max-workers');
+  const defaultMaxWorkers = Math.min(4, Math.max(1, Math.floor(cpus().length / 2)));
+  let maxWorkers: number;
+  let maxWorkersSource: 'flag' | 'registry' | 'default';
+  if (maxWorkersFlagRaw !== undefined) {
+    // fix-round 1, AM-1: an invalid `--max-workers` value (0, negative, fractional, `NaN`,
+    // non-numeric) is a USAGE ERROR (exit 2) — mirroring the `--only` unknown-id fail() style —
+    // never a silent fallback to the default. A silent fallback would make a typo'd flag run
+    // uncapped-by-mistake while LOOKING capped (the source line still says "(default)").
+    const maxWorkersOpt = Number(maxWorkersFlagRaw);
+    if (!Number.isInteger(maxWorkersOpt) || maxWorkersOpt < 1) {
+      return fail(`--max-workers must be a positive integer, got '${maxWorkersFlagRaw}'`);
+    }
+    maxWorkers = maxWorkersOpt;
+    maxWorkersSource = 'flag';
+  } else if (parsed.registry.maxWorkers !== undefined) {
+    maxWorkers = parsed.registry.maxWorkers;
+    maxWorkersSource = 'registry';
+  } else {
+    maxWorkers = defaultMaxWorkers;
+    maxWorkersSource = 'default';
+  }
+  // FR-3: inject the ceiling into the command ONLY when it is (detectably) a vitest run and does
+  // not already name the flag itself — an arbitrary testCommand cannot be assumed to accept
+  // `--maxWorkers`. VITEST_MAX_WORKERS is set in the env unconditionally (below, at spawn time)
+  // regardless of this detection, so a vitest command reached indirectly (e.g. through a package
+  // script) is still capped.
+  // fix-round 1, AM-3: detection AND injection are scoped to the VITEST SEGMENT — from the first
+  // `vitest run` token to the next `&&`/`||`/`;`/`|` (or end of string) — and the flag lands right
+  // after `vitest run`, never appended to the tail of a whole (possibly compound) command. A
+  // raw-substring append over the FULL command turned `vitest run … && cleanup` into
+  // `vitest run … && cleanup --maxWorkers=2` (silently handed to `cleanup`, not vitest), and the
+  // existing-flag check could be suppressed by an unrelated `--maxWorkers` substring living outside
+  // the vitest segment entirely (e.g. inside `cleanup`'s own args, or before `vitest run` in the
+  // same command).
+  const vitestRunIdx = testCmd.indexOf('vitest run');
+  const isVitestCommand = vitestRunIdx !== -1;
+  if (isVitestCommand) {
+    const tailFromRun = testCmd.slice(vitestRunIdx);
+    const terminator = /&&|\|\||;|\|/.exec(tailFromRun);
+    const vitestSegment = terminator !== null ? tailFromRun.slice(0, terminator.index) : tailFromRun;
+    if (!vitestSegment.includes('--maxWorkers')) {
+      const insertAt = vitestRunIdx + 'vitest run'.length;
+      testCmd = `${testCmd.slice(0, insertAt)} --maxWorkers=${maxWorkers}${testCmd.slice(insertAt)}`;
+    }
+  }
+  if (!json) {
+    write(isVitestCommand
+      ? `mutation-gate: workers: ${maxWorkers} (${maxWorkersSource})`
+      : 'mutation-gate: workers: n/a — test command is not vitest');
+  }
 
   // Route-b guard mode: `per-entry` (default, strongest — each red entry re-baselines the restored
   // tree, so a flaky neighbour flips THAT entry to INCONCLUSIVE) or `final` (cheap — one re-run at
@@ -12715,22 +13438,29 @@ function cmdMutationGate(
       phase: 'baseline' | 'mutation' | 'rebaseline' | 'final-rebaseline',
       entryId?: string,
     ): MutationGateRunnerObservation => {
+      // FR-3: VITEST_MAX_WORKERS is set in any case — regardless of whether the command was
+      // recognised as vitest and got the `--maxWorkers=<n>` flag injected — so a vitest command
+      // reached indirectly (a wrapper script) is still capped.
+      const extraEnv: Readonly<Record<string, string>> = { VITEST_MAX_WORKERS: String(maxWorkers) };
       if (injectedRunner !== undefined) {
         return injectedRunner(suiteCommand, {
           cwd: copyDir,
           timeoutMs: timeout,
           phase,
+          env: extraEnv,
           ...(entryId !== undefined ? { entryId } : {}),
         });
       }
+      const startedAt = performance.now();
       const run = spawnSync(suiteCommand, {
         cwd: copyDir,
         shell: true,
         encoding: 'utf-8',
         timeout,
         maxBuffer: 64 * 1024 * 1024,
-        env: { ...process.env, FORCE_COLOR: '0' },
+        env: { ...process.env, FORCE_COLOR: '0', ...extraEnv },
       });
+      const elapsedMs = Math.round(performance.now() - startedAt);
       const errorCode = run.error && 'code' in run.error && typeof run.error.code === 'string'
         ? run.error.code
         : undefined;
@@ -12741,10 +13471,28 @@ function cmdMutationGate(
         throw run.error;
       }
       const signal = typeof run.signal === 'string' ? run.signal : undefined;
+      // mutation-gate-timeout-verdict FR-1/FR-2: a genuine ETIMEDOUT must NEVER be read as a real
+      // suite verdict, even when the killed child intercepted the kill signal and exited with its
+      // OWN status (e.g. a SIGTERM handler calling `process.exit(1)`) — MEASURED: node still sets
+      // `run.error.code === 'ETIMEDOUT'` in that case, but a numeric `run.status` used to win the
+      // `typeof run.status !== 'number'` check below, so the timeout was silently reported as
+      // "baseline suite RED (exit 1)", indistinguishable from a real red suite. ETIMEDOUT now short
+      // -circuits to `exitCode: null` unconditionally, before that check runs.
+      if (errorCode === 'ETIMEDOUT') {
+        // Codex round-1 (2026-09-14): report BOTH facts when both exist — a trapped SIGTERM that
+        // exits 1 shows `child exit 1; signal SIGTERM`, a plain kill shows `no exit code; signal …`.
+        const exitPart = typeof run.status === 'number' ? `child exit ${run.status}` : 'child produced no exit code';
+        const childExit = signal === undefined ? exitPart : `${exitPart}; signal ${signal}`;
+        const suggestedMs = Math.max(timeout + 1, Math.ceil(elapsedMs * 2));
+        return {
+          exitCode: null,
+          output: `${String(run.stdout ?? '')}\n${String(run.stderr ?? '')}`,
+          failureReason: `timeout after ${timeout}ms (elapsed ${elapsedMs}ms); ${childExit}; try --timeout ${suggestedMs} or the registry's timeoutMs field`,
+        };
+      }
       let failureReason: string | undefined;
       if (typeof run.status !== 'number') {
-        if (errorCode === 'ETIMEDOUT') failureReason = `timeout after ${timeout}ms${signal === undefined ? '' : `; signal=${signal}`}`;
-        else if (errorCode === 'ENOBUFS') failureReason = 'maxBuffer exceeded (ENOBUFS; 67108864-byte output ceiling)';
+        if (errorCode === 'ENOBUFS') failureReason = 'maxBuffer exceeded (ENOBUFS; 67108864-byte output ceiling)';
         else if (signal !== undefined) failureReason = `child killed by signal ${signal}`;
         else if (errorCode !== undefined) failureReason = `spawn failure code ${errorCode}`;
         else failureReason = 'spawn failure with no error code or signal';
@@ -12996,7 +13744,7 @@ function cmdMutationGate(
 
   const exitCode = mutationGateExitCode(results, baseline.ok);
   if (json) {
-    write(JSON.stringify({ packageDir: pkgDir, registryPath, testCommand: testCmd, rebaselineMode, baseline, results, summary: summarizeMutationResults(results), warnings, internalRetries, exitCode }, null, 2));
+    write(JSON.stringify({ packageDir: pkgDir, registryPath, testCommand: testCmd, rebaselineMode, baseline, results, summary: summarizeMutationResults(results), warnings, internalRetries, ...(selectionMeta !== null ? { selection: selectionMeta } : {}), exitCode }, null, 2));
     return exitCode;
   }
   write(renderMutationReport(results, baseline, pkgDir));
@@ -20144,11 +20892,11 @@ export async function runCli(argv: string[], io: CliIo = {}): Promise<number> {
       case 'auto-canonicalize':
         return await cmdAutoCanonicalize(options, cwd, write);
       case 'publish':
-        return cmdPublish(options, flags, cwd, write, io.publishMirrorRunner, io.publishSiblingDriftFetcher, io.publishPackedInstallRunner, io.publishExecRunner);
+        return cmdPublish(options, flags, cwd, write, io.publishMirrorRunner, io.publishSiblingDriftFetcher, io.publishPackedInstallRunner, io.publishExecRunner, io.publishGateAuditFsLayer, io.publishNpmPackRunner);
       case 'release':
         return cmdRelease(options, flags, cwd, write, io.releaseRunner);
       case 'parity':
-        return cmdParity(options, flags, write, writeErr, cwd);
+        return await cmdParity(options, flags, write, writeErr, cwd);
       case 'registry':
         return cmdRegistry(options, cwd, write);
       case 'benchmark':