create-cmp-cli 0.18.0 → 0.20.0

package/packages/harness/src/lib/steps-cmp.mjs

@@ -0,0 +1,1284 @@
+// steps-cmp.mjs — the Compose Multiplatform STEP PACK (evidence-economics S8b).
+//
+// Every step the lane runs against a create-cmp app, behind one factory. The
+// SPINE — argument parsing, the subprocess helper with its deadline, the
+// markers, the runner (qa/lib/lane-runner.mjs), the receipt, the journal —
+// lives in qa/verify.mjs and knows nothing about Gradle, adb, Maestro or
+// composeApp/. THIS file knows nothing about receipts or argv. A Kotlin
+// backend supplies a different pack to the same spine; that is the split
+// payment-blueprint had to re-implement by hand (2,769 lines) because the two
+// were one file.
+//
+// `ctx` is everything the steps borrow from the spine, passed explicitly so
+// the borrowing is visible: the project root, the Gradle wrapper, the
+// `--rerun` suffix, the run's mode flags, the subprocess helpers (sh reads the
+// running step's deadline; shGradle adds the KSP self-heal), git helpers, and
+// the DEGRADED_PATHS list the receipt reports.
+//
+// The bodies below are the lane's own steps, moved verbatim from verify.mjs.
+
+import fs from "node:fs";
+import path from "node:path";
+import { compareTokenDrift } from "./token-drift.mjs";
+import { evaluateApprovalsGate } from "./approvals.mjs";
+import { TIERS_SATISFYING, clauseTierCoverage, scanCitations, scanSpecClauses, walkFiles } from "./spec-coverage.mjs";
+import { evaluateComponentStoryParity } from "./component-stories.mjs";
+import { evaluateReachability } from "./reachability.mjs";
+import { memoizeStep } from "./step-cache.mjs";
+import { changedWorkingTreePaths, deriveAffectedFilter } from "./affected-tests.mjs";
+import { acquireDeviceLease, releaseDeviceLease, formatHolder } from "./device-lease.mjs";
+import { ARCH_DOC_REL_PATH, SECTION_IDS, regenerateArchDoc } from "./arch-doc.mjs";
+import { DETERMINISM_TIMEZONES, compareOutcomes, parseJUnitOutcomes } from "./determinism.mjs";
+import { evaluateAuditCadence } from "./audit-cadence.mjs";
+import { androidChecksOutcome } from "./step-outcomes.mjs";
+import { checkHarnessIntegrity, describeIntegrity, LOCK_PATH } from "./harness-lock.mjs";
+
+/**
+ * @param {object} ctx
+ * @param {string} ctx.ROOT project root
+ * @param {string} ctx.HERE the qa/ directory
+ * @param {string} ctx.GRADLEW the wrapper invocation
+ * @param {string} ctx.RERUN " --rerun" in full mode, "" in fast
+ * @param {boolean} ctx.fast
+ * @param {boolean} ctx.determinism
+ * @param {string} ctx.profile
+ * @param {"full"|"fast"} ctx.mode
+ * @param {Function} ctx.sh subprocess helper (throws StepTimeout past the step's deadline)
+ * @param {Function} ctx.shGradle sh + the KSP-collision self-heal
+ * @param {Function} ctx.tryGit
+ * @param {Function} ctx.tryGitLines
+ * @param {string[]} ctx.DEGRADED_PATHS degraded-path activations the receipt reports
+ * @returns {{stepsForProfile: Record<string, Function[]>, DEVICE_STEPS: string[],
+ *   FAST_EXCLUDED_NAMES: string[], STEP_FN_BY_NAME: Record<string, Function>,
+ *   stepDeterminism: Function, releaseLease: () => void}}
+ */
+export function createCmpSteps(ctx) {
+  const { ROOT, HERE, GRADLEW, RERUN, fast, determinism, profile, mode, sh, shGradle, tryGit, tryGitLines, DEGRADED_PATHS } = ctx;
+
+// Recursive: desktopTest writes TEST-*.xml flat, but connected (instrumented) results
+// land one directory level down per device (build/outputs/androidTest-results/connected/
+// debug/<device>/TEST-*.xml) — both shapes are summarized by the same walk.
+function junitSummary(dir) {
+  if (!fs.existsSync(dir)) return null;
+  let tests = 0, failures = 0, errors = 0, skipped = 0;
+  const walk = (d) => {
+    for (const entry of fs.readdirSync(d, { withFileTypes: true })) {
+      const p = path.join(d, entry.name);
+      if (entry.isDirectory()) {
+        walk(p);
+        continue;
+      }
+      if (!entry.name.startsWith("TEST-") || !entry.name.endsWith(".xml")) continue;
+      const xml = fs.readFileSync(p, "utf8");
+      const m = xml.match(/<testsuite[^>]*tests="(\d+)"[^>]*skipped="(\d+)"[^>]*failures="(\d+)"[^>]*errors="(\d+)"/);
+      if (m) {
+        tests += Number(m[1]);
+        skipped += Number(m[2]);
+        failures += Number(m[3]);
+        errors += Number(m[4]);
+      }
+    }
+  };
+  walk(dir);
+  return { tests, failures, errors, skipped };
+}
+
+function deviceAttached() {
+  const res = sh("adb devices", { timeout: 10_000 });
+  if (!res.ok) return false;
+  return res.out.split("\n").slice(1).some((l) => /\tdevice$/.test(l.trim().replace(/\s+/g, "\t")));
+}
+
+// ── The machine-global device lease (qa/lib/device-lease.mjs) ───────────────
+// LANE_MARKER above is per-PROJECT; the device is machine-GLOBAL. A scratch app
+// in /tmp and the real app each stamp their own marker and still share the one
+// emulator — nothing stopped two lanes (or a lane and a live console session)
+// driving it at once, which is the observed wedged-adbd / `device offline` /
+// crossed-app-state failure class. Every device-touching step below takes the
+// lease before touching the device.
+//
+// SCOPE DECISION — once per run, not per step: the lease is acquired lazily by
+// the FIRST device step that actually reaches the device and held until the
+// lane exits (released in the same `finally` as LANE_MARKER). A single run must
+// not thrash acquire/release between adjacent device steps, and holding through
+// the desktop steps interleaved among them (a11y sits between tokenDrift and
+// e2eSmoke) costs nothing — nothing else should drive the device mid-lane
+// anyway, which is the whole point.
+//
+// ON CONTENTION THE STEP RETURNS SKIP — NEVER FAIL: nothing is broken; another
+// run legitimately holds the device. This composes with the evidence ladder
+// (qa/lib/evidence-level.mjs): a SKIPped device step simply does not buy its
+// rung, so contention visibly DEGRADES the evidence level (L2 falls back to L1)
+// instead of corrupting the run with a false red — that degradation being
+// honest and visible is exactly why SKIP is the right verdict.
+let laneDeviceLease = null;
+
+/** Serials of devices currently in `device` state (same parse as deviceAttached). */
+function attachedDeviceSerials() {
+  const res = sh("adb devices", { timeout: 10_000 });
+  if (!res.ok) return [];
+  return res.out
+    .split("\n")
+    .slice(1)
+    .map((l) => l.trim())
+    .filter(Boolean)
+    .map((l) => l.split(/\s+/))
+    .filter(([, state]) => state === "device")
+    .map(([serial]) => serial);
+}
+
+/**
+ * Acquire (or confirm) the lane's device lease for a device step.
+ * Returns null when the lane holds the device; otherwise the SKIP result the
+ * step should return verbatim. The serial leased is the one the lane will
+ * actually drive: the single attached device, or ANDROID_SERIAL when several
+ * are attached (adb/Gradle/Maestro honor the same variable). Ambiguity is
+ * SKIPped by name — leasing a guess would protect the wrong device.
+ */
+function leaseDeviceForStep(stepName) {
+  if (laneDeviceLease) return null; // already held for this run
+  const serials = attachedDeviceSerials();
+  if (serials.length === 0) return null; // each step's own guard SKIPs "no device" with its precise reason
+  let serial = serials[0];
+  if (serials.length > 1) {
+    const chosen = process.env.ANDROID_SERIAL;
+    if (chosen && serials.includes(chosen)) {
+      serial = chosen;
+    } else {
+      return {
+        name: stepName,
+        verdict: "SKIP",
+        reason: `${serials.length} devices attached (${serials.join(", ")}) — the lane cannot tell which one it would drive, so it leases none rather than guessing. Set ANDROID_SERIAL to the device this lane should own, or detach the extras.`,
+        durationMs: 0,
+      };
+    }
+  }
+  const res = acquireDeviceLease({ serial, holder: `verify lane ${stepName}`, root: ROOT });
+  if (!res.ok) {
+    return {
+      name: stepName,
+      verdict: "SKIP",
+      reason: `device ${serial} is held by ${formatHolder(res.heldBy)} — device evidence is batched, not concurrent; wait for it or run once when it finishes`,
+      durationMs: 0,
+    };
+  }
+  if (res.reclaimed) {
+    console.error(`· reclaimed a dead device lease on ${serial} (was ${formatHolder(res.reclaimed)})`);
+  }
+  laneDeviceLease = res.handle;
+  return null;
+}
+
+// Settle adb before handing the device to whatever drives it next (Maestro, the
+// instrumented runner). An install task returning 0 means the package manager accepted
+// the APK — NOT that the device is ready to be driven: a reinstall over a running app
+// briefly drops the emulator's adb transport. `adb devices` still says `device`, but a
+// fresh adb client (Maestro's dadb, Gradle's ddmlib) gets `device offline` and dies
+// before the first assertion (observed 4/4 when the live-inspector tier ran earlier in
+// the lane — its port-forward traffic widens the window — and 0/4 when it was skipped).
+// wait-for-device blocks only while the transport is actually down; the kill/start pair
+// ahead of it clears a stale server-side transport entry that survives the device coming
+// back. Neither weakens any assertion — every downstream check still passes on its own
+// merits.
+function settleAdb() {
+  sh("adb kill-server");
+  sh("adb start-server");
+  sh("adb wait-for-device");
+}
+
+// ── Steps ──────────────────────────────────────────────────────────────────
+// Each returns { name, verdict, reason?, durationMs, details? }. Failure
+// reasons are worded for an AI collaborator to act on.
+
+// Spec ↔ test drift gate — pure Node, no Gradle. The clause/citation scan
+// itself lives in qa/lib/spec-coverage.mjs — the SAME scan feature-brief.mjs
+// derives doneness from, so this gate and the Features view can never disagree
+// about a clause. This step owns only the orphan decision + bookkeeping.
+// The first question any verdict depends on: is the lane that is about to
+// issue it the lane this app was given?
+//
+// Without this the receipt is unfalsifiable in one specific way — edit
+// qa/verify.mjs to force every step PASS and the receipt still validates,
+// because the edited file is simply part of the hashed input surface. Hashing
+// the machine-owned region against qa/harness.lock.json closes that: the lane
+// cannot vouch for itself while modified.
+//
+// DELIBERATELY NOT MEMOIZED. Every other pure-Node step can serve a cached
+// PASS when its inputs are unchanged; a cached PASS on an integrity check is
+// precisely the failure it exists to prevent, and 34 file reads are too cheap
+// to be worth the risk.
+//
+// Three states, three verdicts:
+//   intact    PASS
+//   modified  FAIL — named files, with the command that restores them
+//   unlocked  SKIP — an app stamped before locks existed. Nothing is known to
+//             be wrong, but nothing is proven either; recording the gap keeps
+//             the pipeline honest instead of quietly passing.
+function stepHarnessIntegrity() {
+  const started = Date.now();
+  const r = checkHarnessIntegrity(ROOT);
+  const base = { name: "harnessIntegrity", durationMs: Date.now() - started, harness: r };
+
+  if (r.status === "intact") {
+    return { ...base, verdict: "PASS", note: describeIntegrity(r) };
+  }
+  if (r.status === "unlocked") {
+    return {
+      ...base,
+      verdict: "SKIP",
+      reason: `no ${LOCK_PATH} — this app was stamped before harness locks existed. ` +
+        "`npx create-cmp-cli upgrade --harness` records one.",
+    };
+  }
+
+  const named = [
+    ...r.modified.map((f) => `modified  ${f}`),
+    ...r.missing.map((f) => `missing   ${f}`),
+    ...r.extra.map((f) => `unrecorded ${f}`),
+  ];
+  return {
+    ...base,
+    verdict: "FAIL",
+    reason:
+      `the verify lane has been modified since it was installed — ${describeIntegrity(r)}. ` +
+      "Lane code is machine-owned: it is byte-identical in every create-cmp app and carries " +
+      "no app content, so a local edit is either an accident, a half-applied upgrade, or an " +
+      "attempt to make this receipt say something the lane would not. Restore it with " +
+      "`npx create-cmp-cli upgrade --harness`, which also reports any genuine local patch " +
+      "instead of discarding it.",
+    files: named,
+  };
+}
+
+function stepSpecCoverage() {
+  const started = Date.now();
+  const specsDir = path.join(ROOT, "specs");
+  if (!fs.existsSync(specsDir)) {
+    return { name: "specCoverage", verdict: "SKIP", reason: "no specs/ directory in this project", durationMs: Date.now() - started };
+  }
+
+  const clauses = scanSpecClauses(ROOT);
+  const tags = scanCitations(ROOT);
+  const searchDirs = [path.join(ROOT, "composeApp/src"), path.join(ROOT, "qa/e2e")];
+  const files = searchDirs.flatMap((d) => walkFiles(d, [".kt", ".kts", ".yaml", ".yml"]));
+
+  const citedIds = new Set(tags.map((t) => t.id));
+  const orphanClauses = [...clauses.entries()].filter(([, c]) => !c.withdrawn).filter(([id]) => !citedIds.has(id));
+  const orphanTags = tags.filter((t) => !clauses.has(t.id) || clauses.get(t.id).withdrawn);
+
+  const tiers = clauseTierCoverage(clauses, tags);
+
+  if (orphanClauses.length === 0 && orphanTags.length === 0 && tiers.unmetTier.length === 0) {
+    // Tier visibility, still not a gate for UNDECLARED clauses (industry rule:
+    // instrument before you police). A clause cited only from desktop-tier tests can
+    // still hide a platform-behavior bug — both production apps shipped
+    // alarm/notification defects behind clauses that were "covered" by JVM tests
+    // androidMain never ran under. The line names them; the instrumented seam
+    // (androidChecks) is where such clauses earn a citation.
+    //
+    // A clause that DECLARED `[tier: …]` is policed above — that is the second move
+    // this note's first move was always waiting for.
+    return {
+      name: "specCoverage",
+      verdict: "PASS",
+      durationMs: Date.now() - started,
+      details: {
+        clauses: [...clauses.values()].filter((c) => !c.withdrawn).length,
+        withdrawn: [...clauses.values()].filter((c) => c.withdrawn).length,
+        tags: tags.length,
+        files: files.length,
+        tierNote: tiers.summaryLine,
+      },
+    };
+  }
+
+  const lines = ["Spec coverage broken — the spec and the tests have drifted apart:"];
+  // The competence check, first: an existing-but-blind citation is a subtler
+  // failure than a missing one, and its message has to say WHY the citation it
+  // can see does not count.
+  for (const u of tiers.unmetTier) {
+    const has = u.tiers.length ? `cited only from ${u.tiers.join(", ")}` : "cited by nothing";
+    lines.push(
+      `  [${u.id}] ${u.file} — declares [tier: ${u.requiredTier}] but is ${has}. ` +
+        `A test on those tiers cannot observe this promise (no process lifecycle, no OS facts, no real device). ` +
+        `Add a citing test in ${TIERS_SATISFYING[u.requiredTier].join(" or ")} — and note that tier SKIPPING for want of a device ` +
+        `leaves this clause unproven, which is the point: "I could not check this" is a failure, not a quieter rung.`,
+    );
+  }
+  for (const [id, c] of orphanClauses) {
+    lines.push(`  [${id}] ${c.file} — no durable test cites this clause. Write the test (tag it '// SPEC: ${id}') or withdraw the clause (strike it through).`);
+  }
+  for (const t of orphanTags) {
+    const known = clauses.get(t.id);
+    if (known?.withdrawn) {
+      lines.push(`  // SPEC: ${t.id} at ${t.file}:${t.line} — the test verifies withdrawn behavior (clause ${t.id} in ${known.file} is struck through). Remove the test or un-withdraw the clause.`);
+    } else {
+      lines.push(`  // SPEC: ${t.id} at ${t.file}:${t.line} — no such clause in specs/. Add the clause (AI proposes, human confirms) or fix the id.`);
+    }
+  }
+
+  return {
+    name: "specCoverage",
+    verdict: "FAIL",
+    reason: lines.join("\n"),
+    durationMs: Date.now() - started,
+    details: {
+      clauses: [...clauses.values()].filter((c) => !c.withdrawn).length,
+      withdrawn: [...clauses.values()].filter((c) => c.withdrawn).length,
+      tags: tags.length,
+      files: files.length,
+    },
+  };
+}
+
+// Human-approval gate (VERIFICATION-LAYER-DESIGN.md §2) — pure Node, no Gradle,
+// same grouping as specCoverage. The decision itself lives in
+// qa/lib/approvals.mjs (evaluateApprovalsGate); this step only adds the
+// name/duration bookkeeping every step in this file carries.
+function stepApprovals() {
+  const started = Date.now();
+  const { verdict, reason, statuses } = evaluateApprovalsGate(ROOT);
+  return {
+    name: "approvals",
+    verdict,
+    reason,
+    durationMs: Date.now() - started,
+    details: { artifacts: statuses.map((s) => ({ id: s.id, status: s.status, hash: s.hash })) },
+  };
+}
+
+// There is deliberately NO feature-doneness step here (CHANGE-FLOW-DESIGN.md
+// §7): a feature's doneness is DERIVED from gates this lane already runs —
+// specCoverage fails an uncited clause, the test steps fail a broken promise,
+// and the receipt's inputs.hash attests the tree. A second mechanism would be
+// a second truth.
+
+// Component ↔ story parity gate (STUDIO-REDESIGN.md §3.3) — pure Node, no
+// Gradle, same grouping as specCoverage/approvals. The decision itself lives
+// in qa/lib/component-stories.mjs (evaluateComponentStoryParity); this step
+// only adds the name/duration bookkeeping every step in this file carries.
+function stepComponentStories() {
+  const started = Date.now();
+  const { verdict, reason, details } = evaluateComponentStoryParity(ROOT);
+  return { name: "componentStories", verdict, reason, durationMs: Date.now() - started, details };
+}
+
+// Navigation-reachability gate (task FI-7, docs/AUTONOMY-GAPS.md §3) — pure
+// Node, no Gradle, same grouping as specCoverage/approvals/componentStories.
+// The decision itself lives in qa/lib/reachability.mjs (evaluateReachability);
+// this step only adds the name/duration bookkeeping every step in this file
+// carries. Closes the exact hole a real feature slipped through: every other
+// gate PASSed while its screen was wired into nothing.
+function stepReachability() {
+  const started = Date.now();
+  const { verdict, reason, details } = evaluateReachability(ROOT);
+  return { name: "reachability", verdict, reason, durationMs: Date.now() - started, details };
+}
+
+// Architecture-doc freshness gate (Wave B, docs/proposals/architecture-document-
+// standard.md §6) — pure Node, no Gradle, same grouping as specCoverage/
+// approvals. The decision itself lives in qa/lib/arch-doc.mjs
+// (regenerateArchDoc); this step only adds the name/duration bookkeeping every
+// step in this file carries, plus wording the FAIL reason for an AI
+// collaborator (name the stale/missing section, name the fix command).
+function stepArchDoc() {
+  const started = Date.now();
+  const elapsed = () => Date.now() - started;
+
+  const result = regenerateArchDoc(ROOT);
+  if (!result.ok) {
+    return { name: "archDoc", verdict: "SKIP", reason: `${result.reason} — nothing to check`, durationMs: elapsed() };
+  }
+  if (result.unknownSections.length > 0) {
+    return {
+      name: "archDoc",
+      verdict: "FAIL",
+      reason: `${ARCH_DOC_REL_PATH} has cmp:generated marker(s) with no registered generator: ${result.unknownSections.join(", ")} — add a generator in qa/lib/arch-doc.mjs or remove the marker.`,
+      durationMs: elapsed(),
+    };
+  }
+
+  const stale = result.changed || result.missingSections.length > 0;
+  if (!stale) {
+    return { name: "archDoc", verdict: "PASS", durationMs: elapsed(), details: { sectionsChecked: SECTION_IDS.length } };
+  }
+
+  const lines = [`${ARCH_DOC_REL_PATH} is stale — a generated section no longer matches the tree:`];
+  for (const id of result.changedSections) {
+    lines.push(`  [${id}] regenerating would change this section.`);
+  }
+  for (const id of result.missingSections) {
+    lines.push(`  [${id}] marker missing from the doc entirely — never generated.`);
+  }
+  lines.push("Run: node qa/arch-doc.mjs");
+  return {
+    name: "archDoc",
+    verdict: "FAIL",
+    reason: lines.join("\n"),
+    durationMs: elapsed(),
+    details: { changedSections: result.changedSections, missingSections: result.missingSections },
+  };
+}
+
+// Schema-history gate — pure Node + git, no Gradle, same grouping as the other
+// evidence checks. Room's exportSchema writes one <version>.json per database per
+// target under composeApp/schemas/. Every version EXCEPT the current highest is a
+// frozen historical record of a database that shipped: migrations are written and
+// validated against those exact bytes, so a regeneration that rewrites them
+// silently corrupts the baseline every future migration is proven against. Only
+// the highest version is the live, in-progress schema — free to change or appear
+// (that IS the current change). This gate exists because schema regeneration
+// looks like harmless build output right up until a shipped user's upgrade fails.
+function stepSchemaHistory() {
+  const started = Date.now();
+  const elapsed = () => Date.now() - started;
+  const schemasRel = path.join("composeApp", "schemas");
+  const schemasRoot = path.join(ROOT, schemasRel);
+
+  if (!fs.existsSync(schemasRoot)) {
+    return { name: "schemaHistory", verdict: "SKIP", reason: "no exported Room schemas (composeApp/schemas/ absent) — nothing frozen to guard", durationMs: elapsed() };
+  }
+  const gitTop = tryGit("rev-parse --show-toplevel");
+  if (!gitTop || !tryGit("rev-parse HEAD")) {
+    return { name: "schemaHistory", verdict: "SKIP", reason: "no git history yet — schema versions have no committed baseline to be frozen against", durationMs: elapsed() };
+  }
+
+  // Every directory holding versioned schema JSONs, with its highest version on disk.
+  const versionFile = /^(\d+)\.json$/;
+  const maxVersionByDir = new Map(); // absolute dir path -> highest N among its N.json files
+  const walkSchemas = (dir) => {
+    for (const entry of fs.readdirSync(dir, { withFileTypes: true })) {
+      const p = path.join(dir, entry.name);
+      if (entry.isDirectory()) walkSchemas(p);
+      else {
+        const m = entry.name.match(versionFile);
+        if (m) maxVersionByDir.set(dir, Math.max(maxVersionByDir.get(dir) ?? 0, Number(m[1])));
+      }
+    }
+  };
+  walkSchemas(schemasRoot);
+
+  // Tracked schema files whose committed bytes no longer match the tree (staged or
+  // unstaged; deletions included). Paths come back relative to the git toplevel.
+  // Untracked files never appear here — a brand-new version file is by definition
+  // not yet frozen history.
+  const dirtyFiles = tryGitLines(`diff --name-only HEAD -- "${schemasRel.replace(/\\/g, "/")}"`);
+
+  const violations = [];
+  for (const rel of dirtyFiles) {
+    const abs = path.resolve(gitTop, rel);
+    const m = path.basename(abs).match(versionFile);
+    if (!m) continue; // not a versioned schema JSON
+    const version = Number(m[1]);
+    const dirMax = maxVersionByDir.get(path.dirname(abs));
+    // The highest version currently on disk is the live schema — dirty is fine.
+    // Anything else (a lower version, or a file whose whole directory is gone)
+    // is rewritten/deleted history.
+    if (dirMax !== undefined && version === dirMax) continue;
+    violations.push(rel);
+  }
+
+  if (violations.length === 0) {
+    return { name: "schemaHistory", verdict: "PASS", durationMs: elapsed(), details: { schemaDirs: maxVersionByDir.size } };
+  }
+
+  const lines = [
+    "Historical Room schema files were modified or deleted — these are frozen records of shipped databases, and regeneration must never rewrite them (migrations are validated against these exact bytes). Only the current highest version may change. Restore each file:",
+  ];
+  for (const rel of violations) lines.push(`  git checkout -- ${rel}`);
+  lines.push("If you intended a schema change, bump the database version so a NEW <version>.json is exported instead of overwriting history.");
+  return {
+    name: "schemaHistory",
+    verdict: "FAIL",
+    reason: lines.join("\n"),
+    durationMs: elapsed(),
+    details: { schemaDirs: maxVersionByDir.size, violations },
+  };
+}
+
+function stepBuild() {
+  const res = shGradle(`${GRADLEW} :composeApp:assembleDebug --console=plain`);
+  return {
+    name: "build",
+    verdict: res.ok ? "PASS" : "FAIL",
+    reason: res.ok ? undefined : `assembleDebug failed — fix the build before anything else:\n${res.out.split("\n").filter((l) => /error|FAILURE/i.test(l)).slice(0, 12).join("\n")}`,
+    durationMs: res.durationMs,
+  };
+}
+
+// The build nobody runs until the day they need it.
+//
+// assembleDebug passing says nothing about assembleRelease: R8 and `lintVital` only run on
+// the release variant, and BuildConfig is generated PER BUILD TYPE, so a constant declared
+// in one and not the other is a compile error that only release ever sees. All three of
+// those bit this template at once, and none of them were visible from a green debug lane —
+// the first release build ever attempted (2026-07-29) failed three times over.
+//
+// So release is proven at the checkpoint, not discovered at launch. Unsigned: signing needs
+// a keystore, which belongs to whoever ships the app, and this step is about the shrinker
+// and the build graph rather than the signature.
+function stepReleaseBuild() {
+  const res = shGradle(`${GRADLEW} :composeApp:assembleRelease --console=plain`);
+  return {
+    name: "releaseBuild",
+    verdict: res.ok ? "PASS" : "FAIL",
+    reason: res.ok
+      ? undefined
+      : `assembleRelease failed — the shippable build is broken even though the debug one is fine:\n${res.out
+          .split("\n")
+          .filter((l) => /error|FAILURE|Missing class|Unresolved/i.test(l))
+          .slice(0, 12)
+          .join("\n")}`,
+    durationMs: res.durationMs,
+  };
+}
+
+// Runs a filtered slice of the JVM test tier and names the verdict after the gate it proves.
+// The full suite already ran in unitTests; the filtered slices stay cheap (compilation is
+// cached) while `--rerun` forces the tests themselves to EXECUTE — see stepUnitTests.
+// In fast mode the flag is omitted (RERUN, defined with the mode flags above): the
+// integrity mechanism belongs to the runs that produce integrity-bearing artifacts, and a
+// fast receipt has already declared itself non-evidence.
+function gradleTestStep(name, testsFilter, failHint) {
+  // Named explicitly: a function returned from a factory has no inferred name,
+  // so the runner narrated conformance / goldenTrees / a11y as null and gave
+  // them the default deadline. Same latent class as the memoized wrappers.
+  const step = () => {
+    const res = shGradle(`${GRADLEW} :composeApp:desktopTest${RERUN} --tests "${testsFilter}" --console=plain`);
+    return {
+      name,
+      verdict: res.ok ? "PASS" : "FAIL",
+      reason: res.ok
+        ? undefined
+        : `${failHint}\n${res.out.split("\n").filter((l) => /FAILED|\[(ARCH|SHELL|HOME)-\d+\]|error:/i.test(l)).slice(0, 15).join("\n")}`,
+      durationMs: res.durationMs,
+    };
+  };
+  Object.defineProperty(step, "name", { value: `step${name.charAt(0).toUpperCase()}${name.slice(1)}` });
+  return step;
+}
+
+function stepUnitTests() {
+  // `--rerun` is EVIDENCE INTEGRITY, not pedantry: without it, Gradle's build cache can
+  // restore a PASS recorded against a *different* tree state (deterministic re-scaffolds
+  // produce byte-identical sources, and golden baselines aren't compile inputs), so the
+  // receipt would attest tests that never executed. Compilation stays cached — only the
+  // test execution is forced. Scoped to FULL mode (see RERUN above): the integrity
+  // mechanism belongs to the runs that produce integrity-bearing artifacts, and a fast
+  // receipt is already declared non-evidence.
+  //
+  // Fast mode additionally scopes the suite to tests plausibly affected by the
+  // working-tree change (qa/lib/affected-tests.mjs): changed .kt files map to
+  // `--tests "*<segment>*"` patterns, with a mandatory blast-radius escape hatch (build
+  // files, DI, theme, shared components, qa/, anything outside composeApp/src → full
+  // suite) and fail-open on every uncertain case (no git, unmappable change). FALSE
+  // NEGATIVES ARE ACCEPTABLE HERE AND ONLY HERE: the full, unfiltered suite runs at the
+  // checkpoint (the full lane), where done is actually decided. The filter that ran is
+  // reported in the step's note and recorded in the (fast-only) receipt, so a filtered
+  // run can never be mistaken for the full suite.
+  let note;
+  let testsArgs = "";
+  let affected = null;
+  if (fast) {
+    const changed = changedWorkingTreePaths(ROOT);
+    if (changed === null) {
+      note = "full suite — git unavailable, cannot derive the change (fail open)";
+    } else {
+      const filter = deriveAffectedFilter(changed);
+      if (filter.mode === "filtered") {
+        testsArgs = filter.patterns.map((p) => ` --tests "${p}"`).join("");
+        note = `affected: ${filter.patterns.join(", ")} — ${filter.sourcePaths.length} changed source file(s)`;
+        affected = { patterns: filter.patterns, changedFiles: filter.sourcePaths.length };
+      } else {
+        note = `full suite — ${filter.reason}`;
+      }
+    }
+  }
+  let res = shGradle(`${GRADLEW} :composeApp:desktopTest${RERUN}${testsArgs} --console=plain`);
+  if (fast && testsArgs && !res.ok && /No tests found for given includes/.test(res.out)) {
+    // The heuristic filter matched no test class at all (e.g. a feature with no tests
+    // yet). That is the harness's guess being wrong, not the app — fall back to the
+    // full suite in-lane rather than false-redding on our own filter. (RERUN is empty
+    // here by construction — this branch only exists in fast mode.)
+    const retry = shGradle(`${GRADLEW} :composeApp:desktopTest${RERUN} --console=plain`);
+    retry.durationMs += res.durationMs;
+    res = retry;
+    note = "full suite — the affected-test filter matched no tests (fell back)";
+    affected = null;
+    DEGRADED_PATHS.push("affected-test filter matched no tests — fell back to the full desktopTest suite");
+  }
+  const summary = junitSummary(path.join(ROOT, "composeApp/build/test-results/desktopTest"));
+  let details = summary ?? undefined;
+  if (affected) details = { ...(summary ?? {}), affected };
+  return {
+    name: "unitTests",
+    verdict: res.ok ? "PASS" : "FAIL",
+    reason: res.ok
+      ? undefined
+      : `desktopTest failed (${summary ? `${summary.failures + summary.errors} of ${summary.tests} tests` : "see output"}). Fix the failing behavior — do not delete or weaken tests to pass:\n${res.out.split("\n").filter((l) => /FAILED|error:/i.test(l)).slice(0, 12).join("\n")}`,
+    note,
+    durationMs: res.durationMs,
+    details,
+  };
+}
+
+const stepConformance = gradleTestStep(
+  "conformance",
+  "*ArchitectureConformanceTest",
+  "Architecture conformance violated (specs/app-base.spec.md ARCH clauses). The failing rule names the clause, files, and fix:",
+);
+const stepGoldenTrees = gradleTestStep(
+  "goldenTrees",
+  "*GoldenTreeTest",
+  "Golden-tree drift: a screen's rendered STRUCTURE no longer matches qa/golden/. Unintended → fix your change; intended → regenerate with UPDATE_GOLDEN=1 and declare it:",
+);
+const stepA11y = gradleTestStep(
+  "a11y",
+  "*A11yConformanceTest",
+  "A11y gate failed (SHELL-04): interactive nodes must expose a testTag, text, or contentDescription:",
+);
+
+// ── Determinism probe (roadmap §10 item 8) — opt-in, ci-profile ────────────
+// ARCH-13 statically bans ambient time reads — in APP code. A library the
+// app calls can still read the wall clock, and a golden can still depend on
+// the machine's timezone through a seam the static net cannot see (a
+// ViewModel constructed without its injected clock already caused one
+// overnight golden-tree drift). This probe closes that gap DYNAMICALLY: it
+// runs the JVM test tier twice, under two timezones whose local calendar
+// dates never agree — the offsets are 26 hours apart, so any date-derived
+// value differs between the legs at every instant (see DETERMINISM_TIMEZONES
+// in qa/lib/determinism.mjs for why UTC-12/UTC+14 and not UTC/UTC+14) — and
+// FAILs naming every test whose outcome differs between the legs.
+//
+// Mechanics that carry the honesty:
+//  - TZ reaches the test JVM through the environment: Gradle forwards the
+//    client's environment to the daemon on every build, and test workers
+//    fork from the daemon — so the child env below is inherited all the way
+//    down to the JVM whose default timezone the tests see.
+//  - BOTH legs force --rerun. Without it Gradle would mark the second leg
+//    up-to-date (TZ is not a declared build input) and replay the first
+//    leg's results — the probe would then compare a run against its own
+//    echo and certify a determinism it never tested (the build-cache-replay
+//    lesson, again). The legs use the mode-scoped RERUN like every other
+//    desktopTest invocation — and because --determinism is refused alongside
+//    --fast up front, RERUN is always " --rerun" by the time a leg runs.
+//  - Only verdicts and failure output are compared; durations are never even
+//    parsed (qa/lib/determinism.mjs), so a timing wobble is structurally
+//    unable to trip the probe.
+function stepDeterminism() {
+  const started = Date.now();
+  const elapsed = () => Date.now() - started;
+  if (!determinism) {
+    return {
+      name: "determinism",
+      verdict: "SKIP",
+      reason: "determinism probe is opt-in (it runs the JVM test tier twice) — add --determinism to this lane, or run the probe alone: node qa/verify.mjs --determinism",
+      durationMs: elapsed(),
+    };
+  }
+
+  const resultsDir = path.join(ROOT, "composeApp/build/test-results/desktopTest");
+  const legs = [];
+  for (const { tz, label } of DETERMINISM_TIMEZONES) {
+    const res = shGradle(`${GRADLEW} :composeApp:desktopTest${RERUN} --console=plain`, { env: { ...process.env, TZ: tz } });
+    // Parsed NOW, before the next leg overwrites the same results directory.
+    const outcomes = parseJUnitOutcomes(resultsDir);
+    legs.push({ tz, label, ok: res.ok, outcomes, tail: res.out.split("\n").slice(-8).join("\n") });
+  }
+  const [a, b] = legs;
+  const labelA = `TZ=${a.tz} (${a.label})`;
+  const labelB = `TZ=${b.tz} (${b.label})`;
+  const countA = Object.keys(a.outcomes).length;
+  const countB = Object.keys(b.outcomes).length;
+
+  if (countA === 0 && countB === 0) {
+    // Neither leg produced a single test result: the suite failed before
+    // running anything (build error). The probe measured nothing — that is
+    // a FAIL that says so, never a PASS by absence of differences.
+    return {
+      name: "determinism",
+      verdict: "FAIL",
+      reason: `determinism probe could not execute: desktopTest produced no test results under either timezone — the suite fails before running (fix the build first; this is not a timezone difference):\n${a.tail}`,
+      durationMs: elapsed(),
+    };
+  }
+  if (countA === 0 || countB === 0) {
+    const ran = countA > 0 ? { label: labelA, n: countA } : { label: labelB, n: countB };
+    const empty = countA > 0 ? b : a;
+    return {
+      name: "determinism",
+      verdict: "FAIL",
+      reason: `Nondeterminism under timezone shift: desktopTest ran ${ran.n} test(s) under ${ran.label} but produced no results at all under TZ=${empty.tz} (${empty.label}) — the suite itself dies under that zone:\n${empty.tail}`,
+      durationMs: elapsed(),
+      details: { timezones: DETERMINISM_TIMEZONES.map((t) => t.tz) },
+    };
+  }
+
+  const diffs = compareOutcomes(a.outcomes, b.outcomes, labelA, labelB);
+  if (diffs.length > 0) {
+    const lines = [
+      `Nondeterminism under timezone shift — the same tree produced different outcomes under ${labelA} vs ${labelB}. Something reads ambient time or zone past the ARCH-13 net (a library default, an uninjected clock, a golden that captures "today"):`,
+    ];
+    for (const d of diffs.slice(0, 20)) lines.push(`  [${d.step}] ${d.test} — ${d.detail}`);
+    if (diffs.length > 20) lines.push(`  … and ${diffs.length - 20} more differing test(s)`);
+    return {
+      name: "determinism",
+      verdict: "FAIL",
+      reason: lines.join("\n"),
+      durationMs: elapsed(),
+      details: { timezones: DETERMINISM_TIMEZONES.map((t) => t.tz), diffs: diffs.slice(0, 50) },
+    };
+  }
+
+  const failedIdentically = Object.values(a.outcomes).filter((o) => o.status !== "pass" && o.status !== "skip").length;
+  return {
+    name: "determinism",
+    verdict: "PASS",
+    // Identical red is DETERMINISTIC red: the probe's claim ("no timezone
+    // dependence") holds, and the failing tests already belong to
+    // unitTests/goldenTrees, which fail the lane on their own merits — a
+    // second FAIL here would report the same defect twice under a wrong name.
+    note:
+      failedIdentically > 0
+        ? `${failedIdentically} test(s) failed identically under both timezones — deterministic, but red (the owning test steps report it)`
+        : undefined,
+    durationMs: elapsed(),
+    details: { timezones: DETERMINISM_TIMEZONES.map((t) => t.tz), testsCompared: countA },
+  };
+}
+
+// Live tokenDrift tier (harness M4-D): when a debug app + device are available,
+// fetches the declared catalog and the live semantics tree off the debug-only
+// inspector server (127.0.0.1:9500, see composeApp/src/androidDebug/.../
+// InspectorHttpServer.kt) and runs compareTokenDrift() over them — real runtime
+// drift detection, embedded in the evidence receipt.
+//
+// Infrastructure absence (no device, app not running) is NEVER a FAIL — only
+// actual drift is. curl (via the existing synchronous sh() helper) stands in for
+// an HTTP client here because every step in this lane runs synchronously; a
+// couple of short retries cover the debug app's cold start.
+const INSPECTOR_PORT = 9500;
+
+function curlJson(url, timeoutSec = 5) {
+  const res = sh(`curl -s -m ${timeoutSec} -w "\\n%{http_code}" "${url}"`);
+  if (!res.ok) return { ok: false };
+  const out = res.out;
+  const idx = out.lastIndexOf("\n");
+  const code = (idx >= 0 ? out.slice(idx + 1) : "").trim();
+  const bodyText = idx >= 0 ? out.slice(0, idx) : "";
+  if (code !== "200") return { ok: false };
+  try {
+    return { ok: true, body: JSON.parse(bodyText) };
+  } catch {
+    return { ok: false };
+  }
+}
+
+function pollHealth(port, attempts, delaySec) {
+  let health = curlJson(`http://127.0.0.1:${port}/inspect/health`);
+  for (let tries = 1; !health.ok && tries < attempts; tries += 1) {
+    sh(`sleep ${delaySec}`);
+    health = curlJson(`http://127.0.0.1:${port}/inspect/health`);
+  }
+  return health;
+}
+
+function stepTokenDrift() {
+  const started = Date.now();
+  const elapsed = () => Date.now() - started;
+
+  if (!deviceAttached()) {
+    return {
+      name: "tokenDrift",
+      verdict: "SKIP",
+      reason: "no Android device/emulator attached (adb) — runtime token drift needs the live inspector tier",
+      durationMs: elapsed(),
+    };
+  }
+
+  const unreachable = () => ({
+    name: "tokenDrift",
+    verdict: "SKIP",
+    reason: "inspector endpoint not reachable on :9500 (debug app not running?) — launch the debug build to enable the live tier",
+    durationMs: elapsed(),
+  });
+
+  // Machine-global lease before the first device touch (contention = SKIP).
+  const leaseSkip = leaseDeviceForStep("tokenDrift");
+  if (leaseSkip) return { ...leaseSkip, durationMs: elapsed() };
+
+  sh(`adb forward tcp:${INSPECTOR_PORT} tcp:${INSPECTOR_PORT}`);
+  try {
+    let health = curlJson(`http://127.0.0.1:${INSPECTOR_PORT}/inspect/health`);
+    if (!health.ok) {
+      // Debug app may not be running — try to launch it (best-effort: parse the
+      // applicationId out of the Android build config), then give it a moment
+      // to cold-start before giving up.
+      let applicationId = null;
+      try {
+        const gradle = fs.readFileSync(path.join(ROOT, "composeApp/build.gradle.kts"), "utf8");
+        applicationId = gradle.match(/applicationId\s*=\s*"([^"]+)"/)?.[1] ?? null;
+      } catch {
+        applicationId = null;
+      }
+      if (applicationId) {
+        sh(`adb shell am start -n ${applicationId}/.MainActivity`);
+      }
+      health = pollHealth(INSPECTOR_PORT, 5, 2);
+    }
+    if (!health.ok) return unreachable();
+
+    const designSystem = curlJson(`http://127.0.0.1:${INSPECTOR_PORT}/inspect/design-system`);
+    const tree = curlJson(`http://127.0.0.1:${INSPECTOR_PORT}/inspect/tree`);
+    if (!designSystem.ok || !tree.ok) return unreachable();
+
+    const { checked, drifted } = compareTokenDrift(designSystem.body, tree.body);
+
+    if (drifted.length === 0) {
+      return {
+        name: "tokenDrift",
+        verdict: "PASS",
+        durationMs: elapsed(),
+        details: { checked, drifted: 0 },
+      };
+    }
+
+    const lines = ["Runtime token drift — a component's resolved value contradicts the declared design-system catalog:"];
+    for (const d of drifted) {
+      lines.push(
+        `  [${d.node}] token '${d.token}' (${d.facet}) — expected ${d.expected}, resolved ${d.actual}. Update the component to use the token, or update the catalog if the token itself changed.`,
+      );
+    }
+    return {
+      name: "tokenDrift",
+      verdict: "FAIL",
+      reason: lines.join("\n"),
+      durationMs: elapsed(),
+      details: { checked, drifted },
+    };
+  } finally {
+    sh(`adb forward --remove tcp:${INSPECTOR_PORT}`);
+  }
+}
+
+function maestroAvailable() {
+  return sh("maestro --version", { timeout: 15_000 }).ok;
+}
+
+// The e2e guard trio, shared by every step that drives the smoke flow on a device.
+// Returns null when the harness is fully available, else the SKIP result for [name].
+function maestroGuards(name) {
+  if (!fs.existsSync(path.join(ROOT, "qa/e2e"))) {
+    return { name, verdict: "SKIP", reason: "e2e harness not included in this project (--no-e2e)", durationMs: 0 };
+  }
+  if (!deviceAttached()) {
+    return { name, verdict: "SKIP", reason: "no Android device/emulator attached (adb)", durationMs: 0 };
+  }
+  if (!maestroAvailable()) {
+    return { name, verdict: "SKIP", reason: "maestro CLI not installed — curl -fsSL https://get.maestro.mobile.dev | bash", durationMs: 0 };
+  }
+  return null;
+}
+
+// Drives qa/e2e/smoke.yaml against whatever build is installed, with the device hardened
+// for headless/CI automation. Shared by e2eSmoke (debug APK) and releaseSmoke (release
+// APK) so the hardening and the honesty sweep can never drift apart between variants.
+// Without the hardening, a slow or loaded emulator produces false reds that have nothing
+// to do with the app:
+//  - hide_error_dialogs=1 stops Android popping ANR/crash dialogs (e.g. SystemUI under load)
+//    that steal focus over the app — a Maestro assert would then see only the dialog;
+//  - MAESTRO_DRIVER_STARTUP_TIMEOUT gives the UiAutomator2 driver a generous budget to come
+//    up on a slow emulator (the built-in default gives up too early under load).
+// Both are benign, reversible, and only touch the device while the lane is driving it —
+// hide_error_dialogs is restored to its pre-run value (or deleted, returning the device
+// to its default) in the finally below, on every exit path.
+// hide_error_dialogs suppresses the OS dialog, NEVER the underlying event — so after the
+// run we grep the device log for ANR/crash lines the dialog would have shown, and FAIL on
+// them. The eyes must report what automation stability had to hide.
+function runMaestroSmoke(name, priorDurationMs) {
+  const prevHideErrorDialogs = sh("adb shell settings get global hide_error_dialogs").out.trim();
+  sh("adb shell settings put global hide_error_dialogs 1");
+  sh("adb logcat -c"); // clear so the post-run dump only reflects this run
+  try {
+    const res = sh("maestro test qa/e2e/smoke.yaml", { env: { ...process.env, MAESTRO_DRIVER_STARTUP_TIMEOUT: "120000" } });
+    if (!res.ok) {
+      return {
+        name,
+        verdict: "FAIL",
+        reason: `Maestro smoke failed (flow cites the SHELL spec clauses it proves):\n${res.out.split("\n").slice(-15).join("\n")}`,
+        durationMs: priorDurationMs + res.durationMs,
+      };
+    }
+    const anrDump = sh("adb logcat -d -b system,crash,main");
+    const anrRe = /ANR in |FATAL EXCEPTION/i;
+    if (anrDump.ok && anrRe.test(anrDump.out)) {
+      const anrLines = anrDump.out.split("\n").filter((l) => anrRe.test(l)).slice(0, 10).join("\n");
+      return {
+        name,
+        verdict: "FAIL",
+        reason: `Maestro smoke passed, but the device log shows an ANR/crash during the run (hide_error_dialogs only suppresses the OS dialog, never the underlying event):\n${anrLines}`,
+        durationMs: priorDurationMs + res.durationMs,
+      };
+    }
+    return { name, verdict: "PASS", durationMs: priorDurationMs + res.durationMs };
+  } finally {
+    if (prevHideErrorDialogs && prevHideErrorDialogs !== "null") {
+      sh(`adb shell settings put global hide_error_dialogs ${prevHideErrorDialogs}`);
+    } else {
+      sh("adb shell settings delete global hide_error_dialogs");
+    }
+  }
+}
+
+function stepE2eSmoke() {
+  const guard = maestroGuards("e2eSmoke");
+  if (guard) return guard;
+  // Machine-global lease before the first device touch (contention = SKIP).
+  const leaseSkip = leaseDeviceForStep("e2eSmoke");
+  if (leaseSkip) return leaseSkip;
+  const install = shGradle(`${GRADLEW} :composeApp:installDebug --console=plain`);
+  if (!install.ok) {
+    return { name: "e2eSmoke", verdict: "FAIL", reason: "installDebug failed — the APK could not be installed on the attached device", durationMs: install.durationMs };
+  }
+  settleAdb();
+  return runMaestroSmoke("e2eSmoke", install.durationMs);
+}
+
+// Instrumented behavior tier (composeApp/src/androidInstrumentedTest) — the one step
+// whose evidence crosses the process boundary. Alarms, notification channels,
+// full-screen intents, PendingIntent identity, and audio routing are OS facts:
+// desktopTest is a JVM, golden trees are structure, the conformance suite is static,
+// and the Maestro smoke taps UI without asserting anything about the shade or the
+// alarm table. Nine escaped platform-semantics defects across two real apps trace to
+// exactly this blind spot; the hand-built precursor of this step caught two bugs the
+// week it landed. `connectedDebugAndroidTest` builds, installs, and runs the
+// instrumented suite in the app's real process on the attached device.
+//
+// SKIP (never FAIL) on missing infrastructure — no device, or no instrumented sources
+// yet — mirroring e2eSmoke's stance: absence of the tier is recorded honestly, only
+// broken behavior fails.
+function stepAndroidChecks() {
+  const started = Date.now();
+  const instrumentedDir = path.join(ROOT, "composeApp/src/androidInstrumentedTest");
+  const hasSources = fs.existsSync(instrumentedDir) &&
+    walkFiles(instrumentedDir, [".kt"]).length > 0;
+  if (!hasSources) {
+    return {
+      name: "androidChecks",
+      verdict: "SKIP",
+      reason: "no instrumented tests (composeApp/src/androidInstrumentedTest has no Kotlin sources)",
+      durationMs: Date.now() - started,
+    };
+  }
+  if (!deviceAttached()) {
+    return {
+      name: "androidChecks",
+      verdict: "SKIP",
+      reason: "no Android device/emulator attached (adb) — instrumented behavior needs the real process boundary",
+      durationMs: Date.now() - started,
+    };
+  }
+  // Machine-global lease before the first device touch (contention = SKIP).
+  const leaseSkip = leaseDeviceForStep("androidChecks");
+  if (leaseSkip) return { ...leaseSkip, durationMs: Date.now() - started };
+  // Settle before Gradle's own install+drive: earlier lane steps (tokenDrift's
+  // port-forwards, e2eSmoke's reinstall) can leave the transport stale — see settleAdb.
+  settleAdb();
+  // `--rerun` for the same evidence-integrity reason as stepUnitTests: the receipt must
+  // attest tests that EXECUTED on this tree, never a replayed up-to-date verdict.
+  const res = shGradle(`${GRADLEW} :composeApp:connectedDebugAndroidTest --rerun --console=plain`);
+  const summary = junitSummary(path.join(ROOT, "composeApp/build/outputs/androidTest-results/connected"));
+  // Verdict separated from invocation (qa/lib/step-outcomes.mjs): a run that
+  // executed zero tests has observed nothing and must not accuse the change.
+  const outcome = androidChecksOutcome(res, summary, { gradlew: GRADLEW });
+  return {
+    name: "androidChecks",
+    verdict: outcome.verdict,
+    reason: outcome.reason,
+    durationMs: Date.now() - started,
+    // `executed` rides on the receipt so a reader can tell a red that measured
+    // something from a red that measured nothing. Shape preserved: undefined
+    // when there is no summary AND nothing to add, as before.
+    details: summary ? { ...summary, executed: outcome.executed } : outcome.executed ? undefined : { executed: false },
+  };
+}
+
+// Release-APK smoke — the behavior half of stepReleaseBuild. assembleRelease proves R8
+// and the build graph COMPILE; two real bugs were only findable by *running* the release
+// variant (R8 behavior differs from debug). Installs the release APK and drives the same
+// Maestro smoke flow against it. Ship-time cost by design: this step exists only in the
+// `release` profile, never per-change.
+//
+// Honesty notes, both deliberate:
+//  - A template-fresh app has NO release signingConfig (the keystore belongs to whoever
+//    ships), and an unsigned APK cannot be installed. That is a SKIP naming what to
+//    configure, never a FAIL — a fresh scaffold must not red-bar on a keystore it was
+//    never given.
+//  - This step reinstalls NOTHING afterwards: the release build stays on the device,
+//    which is the honest state ("what is installed is what was last proven"). The next
+//    debug install over it will hit INSTALL_FAILED_UPDATE_INCOMPATIBLE (release and debug
+//    signatures differ) — run `adb uninstall <applicationId>` first; the same applies in
+//    reverse here, so that raw Gradle error is translated into the actionable message.
+function stepReleaseSmoke() {
+  const guard = maestroGuards("releaseSmoke");
+  if (guard) return guard;
+
+  let gradleText = "";
+  try {
+    gradleText = fs.readFileSync(path.join(ROOT, "composeApp/build.gradle.kts"), "utf8");
+  } catch {
+    gradleText = "";
+  }
+  const applicationId = gradleText.match(/applicationId\s*=\s*"([^"]+)"/)?.[1] ?? "<applicationId>";
+  if (!/signingConfig/.test(gradleText)) {
+    return {
+      name: "releaseSmoke",
+      verdict: "SKIP",
+      reason:
+        "release APK is unsigned — no signingConfig in composeApp/build.gradle.kts. To enable the release smoke: create a keystore (keytool -genkeypair), declare android.signingConfigs { create(\"release\") { … } } from a gitignored keystore.properties, and set buildTypes.release.signingConfig. The keystore is yours to keep out of the repo.",
+      durationMs: 0,
+    };
+  }
+
+  // Machine-global lease before the first device touch (contention = SKIP).
+  // After the signing check on purpose: an unsigned template SKIPs on the
+  // keystore without ever needing the device.
+  const leaseSkip = leaseDeviceForStep("releaseSmoke");
+  if (leaseSkip) return leaseSkip;
+
+  const install = shGradle(`${GRADLEW} :composeApp:installRelease --console=plain`);
+  if (!install.ok) {
+    if (/INSTALL_FAILED_UPDATE_INCOMPATIBLE/.test(install.out)) {
+      return {
+        name: "releaseSmoke",
+        verdict: "FAIL",
+        reason: `installRelease refused: the device holds a build with a different signature (usually the debug build from an earlier lane step). Android never installs across signatures — run \`adb uninstall ${applicationId}\` and re-run the release profile. This is a device-state conflict, not a build defect.`,
+        durationMs: install.durationMs,
+      };
+    }
+    if (/SigningConfig|not signed|INSTALL_PARSE_FAILED_NO_CERTIFICATES/i.test(install.out)) {
+      return {
+        name: "releaseSmoke",
+        verdict: "SKIP",
+        reason: "release APK is not installable — signing is not fully configured (see composeApp/build.gradle.kts signingConfigs). Configure a release keystore to enable the release smoke.",
+        durationMs: install.durationMs,
+      };
+    }
+    return {
+      name: "releaseSmoke",
+      verdict: "FAIL",
+      reason: `installRelease failed — the shippable APK could not be installed:\n${install.out.split("\n").filter((l) => /error|FAILURE|INSTALL_/i.test(l)).slice(0, 12).join("\n")}`,
+      durationMs: install.durationMs,
+    };
+  }
+  settleAdb();
+  return runMaestroSmoke("releaseSmoke", install.durationMs);
+}
+
+// ── Audit cadence (roadmap §10 item 9) — a REPORT, never a gate ────────────
+// cmp-audit (the adversarial platform-semantics audit) found six latent
+// defects the first time a human happened to ask for it — which is exactly
+// why it must not depend on someone remembering to ask. This step is the
+// cheapest honest replacement for that memory: at ship time (release
+// profile) the receipt lists which androidMain subsystems changed since
+// their last RECORDED audit (qa/audits.jsonl, appended by
+// node qa/record-audit.mjs). The derivation lives in
+// qa/lib/audit-cadence.mjs; this step adds only the bookkeeping every step
+// carries — and by construction it maps every outcome to PASS or SKIP,
+// never FAIL: audit debt is a judgment call (a rename is not six latent
+// defects), and a gate here would teach people to game the ledger, which
+// would destroy the only value it has.
+function stepAuditCadence() {
+  const started = Date.now();
+  const report = evaluateAuditCadence(ROOT);
+  if (!report.ok) {
+    return { name: "auditCadence", verdict: "SKIP", reason: report.reason, durationMs: Date.now() - started };
+  }
+  return {
+    name: "auditCadence",
+    verdict: "PASS",
+    note: report.summary,
+    durationMs: Date.now() - started,
+    details: {
+      packageRoot: report.packageRoot,
+      subsystems: report.subsystems.map((s) => ({
+        name: s.name,
+        status: s.status,
+        changedFiles: s.changedFiles,
+        lastAudit: s.audit ? { sha: s.audit.sha, at: s.audit.at, by: s.audit.by } : null,
+      })),
+      lines: report.lines,
+    },
+  };
+}
+
+// ── Lane ───────────────────────────────────────────────────────────────────
+
+// Device-dependent steps, in lane order. Used twice: receipt STRENGTH (which
+// on-device steps actually PASSed — see below, where the receipt is built) and
+// the --fast exclusion (with releaseBuild added), so the "device/slow tier"
+// can never mean two different lists.
+const DEVICE_STEPS = ["e2eSmoke", "tokenDrift", "androidChecks", "releaseSmoke"];
+
+// ── Fast-mode memoization of the pure-Node steps (qa/lib/step-cache.mjs) ────
+// These five steps run no Gradle, shell out to nothing, and are pure functions
+// of files on disk — so in FAST mode an unchanged input set reuses the last
+// PASS as verdict "CACHED" (rendered distinctly; only a PASS is ever reused,
+// a cached FAIL/SKIP always re-runs). THE FULL LANE NEVER CONSULTS THE CACHE —
+// deliberately: it keeps the integrity property absolute rather than "absolute
+// unless a cache says otherwise". A full run still WRITES entries so the next
+// fast run benefits. schemaHistory is NOT here even though it runs no Gradle:
+// it shells out to git and its verdict depends on HEAD state, not only file
+// bytes — memoizing it on a content hash could go silently stale.
+//
+// Each input set is the step's ACTUAL read surface, over-declared where cheap
+// (a too-broad set only costs cache misses; a too-narrow one is a
+// silently-stale gate — the worst possible bug here):
+//   specCoverage     reads specs/*.spec.md + citations under composeApp/src
+//                    and qa/e2e (qa/lib/spec-coverage.mjs)
+//   approvals        reads qa/approvals.json + every governed artifact file:
+//                    specs/, docs/features/, docs/ARCHITECTURE.md, and the
+//                    exemplar/theme/components Kotlin under composeApp/src
+//                    (qa/lib/approvals.mjs listGovernedArtifacts)
+//   componentStories reads commonMain presentation/components and desktopMain
+//                    inspector sources — both under composeApp/src
+//   reachability     reads commonMain Kotlin (composeApp/src) + the unrouted
+//                    declarations in docs/features/
+//   archDoc          reads docs/ARCHITECTURE.md, docs/adr/, specs/intent.md
+//                    (over-declared to all of specs/), and every source-set's
+//                    Kotlin under composeApp/src (qa/lib/arch-doc.mjs)
+const MEMOIZED_STEP_INPUTS = {
+  specCoverage: ["specs", "composeApp/src", "qa/e2e"],
+  approvals: ["qa/approvals.json", "specs", "docs/features", "docs/ARCHITECTURE.md", "composeApp/src"],
+  componentStories: ["composeApp/src"],
+  reachability: ["composeApp/src", "docs/features"],
+  archDoc: ["docs/ARCHITECTURE.md", "docs/adr", "specs", "composeApp/src"],
+};
+
+// The wrapper carries the step's NAME explicitly. An inner arrow returned from
+// a factory has no inferred name, so the runner's stepDisplayName() read these
+// as null — the marker narrated "specCoverage" as nothing, and its deadline
+// fell to the 30-minute default. Latent since drive-narration; surfaced by the
+// pack's own test the moment names were asserted at runtime rather than in
+// source.
+const memoized = (stepName, stepFn) => {
+  const wrapped = () => memoizeStep({ fast, root: ROOT, stepName, inputs: MEMOIZED_STEP_INPUTS[stepName], run: stepFn });
+  Object.defineProperty(wrapped, "name", { value: `step${stepName.charAt(0).toUpperCase()}${stepName.slice(1)}Memo` });
+  return wrapped;
+};
+
+const stepSpecCoverageMemo = memoized("specCoverage", stepSpecCoverage);
+const stepApprovalsMemo = memoized("approvals", stepApprovals);
+const stepComponentStoriesMemo = memoized("componentStories", stepComponentStories);
+const stepReachabilityMemo = memoized("reachability", stepReachability);
+const stepArchDocMemo = memoized("archDoc", stepArchDoc);
+
+const stepsForProfile = {
+  // scaffold: what `create-cmp --verify` proves at stamp time — specCoverage,
+  // the full JVM tier (unit + conformance + golden + UI tests) plus the Android build.
+  scaffold: [stepHarnessIntegrity, stepSpecCoverageMemo, stepApprovalsMemo, stepComponentStoriesMemo, stepReachabilityMemo, stepArchDocMemo, stepSchemaHistory, stepBuild, stepUnitTests],
+  // smoke (docs/GATE-RULES.md Rule 0, docs/PRINCIPLES.md #2): the smallest
+  // end-to-end lane — every pure-Node step through the REAL runner, marker,
+  // receipt and journal, and NO Gradle, no device, no network. Its job is to
+  // prove the framework RETURNS, fast, in both directions, before any real
+  // work is pointed at it. scripts/framework-check.mjs drives it: PASS on a
+  // fresh scaffold, then FAIL BY NAME on one planted spec edit, each bounded
+  // in seconds. Its receipt is refused as done-evidence (qa/receipt-check.mjs)
+  // exactly like --fast: it proves the instrument, never the change.
+  smoke: [stepHarnessIntegrity, stepSpecCoverageMemo, stepApprovalsMemo, stepComponentStoriesMemo, stepReachabilityMemo, stepArchDocMemo, stepSchemaHistory],
+  local: [
+    // First, always: every verdict below is only worth what the lane issuing
+    // it is worth.
+    stepHarnessIntegrity,
+    stepSpecCoverageMemo,
+    stepApprovalsMemo,
+    stepComponentStoriesMemo,
+    stepReachabilityMemo,
+    stepArchDocMemo,
+    stepSchemaHistory,
+    stepBuild,
+    stepUnitTests,
+    stepConformance,
+    stepGoldenTrees,
+    stepTokenDrift,
+    stepA11y,
+    // Release stays OUT of `scaffold`: stamp-time --verify promises a green first build, and
+    // an R8 pass would add minutes to every scaffold to re-prove what this step proves here.
+    // local + ci is where release rot gets caught before it reaches anyone.
+    //
+    // And it sits AFTER the cheap tier, not before it. The lane runs every step
+    // regardless of failures, so the order costs nothing on a green run — but
+    // ahead of them, a red unit test was reported only once R8 had finished,
+    // which on a real change is minutes of waiting to be told something the JVM
+    // knew in seconds. Cheap high-signal checks report first.
+    stepReleaseBuild,
+    stepE2eSmoke,
+    // androidChecks joins local BY the file's own convention, not despite it: local's
+    // contract (see USAGE) is "everything; device-dependent steps SKIP when no device is
+    // attached" — device presence is the opt-in, exactly as e2eSmoke and tokenDrift
+    // already work. A developer with no device attached pays nothing here; one who
+    // attached an emulator has already opted into the device tier's cost. Hiding this
+    // step in ci-only would make local's documented contract a lie and re-open the gap
+    // this tier closes (androidMain test-invisible in the profile people actually run).
+    // Last on purpose: the cheap desktop verdicts and the smoke land first.
+    stepAndroidChecks,
+  ],
+};
+// ci = local + the determinism probe's row — the first place ci diverges
+// from local. The probe is OPT-IN (the step SKIPs unless --determinism was
+// passed: it doubles the JVM test tier's cost), but its row lives in the ci
+// profile so a ci receipt always records whether the probe ran — an honest,
+// visible gap beats an invisible one ("SKIPs are recorded so the pipeline
+// stays honest", per the profile's own contract). local deliberately does
+// NOT carry the row: the per-change developer profile is not where a
+// deliberate double-run belongs.
+stepsForProfile.ci = [...stepsForProfile.local, stepDeterminism];
+// release = everything ci proves PLUS the audit-cadence report and the
+// release-APK behavior smoke. The expensive proofs are profile-tiered by
+// decision: per-change stays fast (local/ci pay for the release COMPILE via
+// releaseBuild, already in the set), and the release-variant *behavior* cost
+// lands once, at ship time. auditCadence (a report, never a gate) also
+// belongs to ship time — "what moved in androidMain since its last
+// adversarial audit?" is the question asked before shipping, not per edit.
+// releaseSmoke runs last so the device ends the run holding the exact build
+// that was proven.
+stepsForProfile.release = [...stepsForProfile.ci, stepAuditCadence, stepReleaseSmoke];
+// nightly (evidence-economics S6 / proposal P4): the stage for proofs whose cost
+// scales with the SUITE rather than with the change — the determinism probe
+// today (forced on above; `--determinism` is implied), and the place any
+// future mutation / load / chaos step lands, so the placement decision is made
+// once instead of per expensive step. It proves the harness and the tree's
+// invariants, not a change: qa/receipt-check.mjs refuses its receipt as
+// done-evidence, exactly as it refuses --fast. Same step set as ci on purpose —
+// what differs is what is forced, and what the receipt is allowed to mean.
+stepsForProfile.nightly = [...stepsForProfile.ci];
+
+const FAST_EXCLUDED_NAMES = [...DEVICE_STEPS, "releaseBuild"];
+const STEP_FN_BY_NAME = {
+  e2eSmoke: stepE2eSmoke,
+  tokenDrift: stepTokenDrift,
+  androidChecks: stepAndroidChecks,
+  releaseSmoke: stepReleaseSmoke,
+  releaseBuild: stepReleaseBuild,
+};
+for (const name of FAST_EXCLUDED_NAMES) {
+  if (!STEP_FN_BY_NAME[name]) {
+    // Drift guard: a new device-tier step must be mapped here or --fast would silently run it.
+    console.error(`internal: fast-excluded step "${name}" has no entry in STEP_FN_BY_NAME — fix qa/verify.mjs`);
+    process.exit(2);
+  }
+}
+
+  return {
+    stepsForProfile,
+    DEVICE_STEPS,
+    FAST_EXCLUDED_NAMES,
+    STEP_FN_BY_NAME,
+    stepDeterminism,
+    // The device lease is held to the very end of the run (see the scope
+    // decision above); the spine releases it in the runner's finally.
+    releaseLease: () => {
+      if (laneDeviceLease) releaseDeviceLease(laneDeviceLease);
+    },
+  };
+}