test(e2e): accept managed onboarding outcomes

test/e2e/README.md

@@ -1268,10 +1268,13 @@ No PR E2E controller dispatches the risk plan.
 The `full-e2e` target enforces a separate hard acceptance contract for the
 first fresh onboarding path in that job. It measures from the onboard root span
 (a conservative anchor before wizard step `[1/8]`) through the first non-empty
-agent response, requires the local BuildKit prebuild for the NemoClaw-generated
-context without a gateway-builder fallback, enforces the calibrated root and
-phase limits in the budget file, and limits the longest onboard output gap to
-60 seconds. A violation fails
+agent response and reads the registered workload receipt. A `legacy-dockerfile`
+receipt requires the local BuildKit prebuild without a gateway-builder fallback.
+A `managed-image` receipt instead requires an exact digest that matches the
+registered sandbox image tag, a non-empty publication cohort, and an exact
+40-character source revision, and it forbids a local BuildKit prebuild. Both
+paths enforce the calibrated root and phase limits in the budget file and limit
+the longest onboard output gap to 60 seconds. A violation fails
 `full-e2e`, and the target writes its evidence to `onboard-progress-budget.json`.
 The artifact records the first-turn command wall clock and OpenClaw's internal
 agent duration separately. Older or malformed OpenClaw output records an

test/e2e/live/full-e2e-workload-evidence.ts

@@ -0,0 +1,39 @@
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+import { readManagedWorkloadAuthority } from "../../../src/lib/onboard/workload/authority.ts";
+import { load as loadSandboxRegistry } from "../../../src/lib/state/registry/persistence.ts";
+
+export function readFullE2eColdWorkloadEvidence(
+  sandboxName: string,
+  usedBuildKitPrebuild: boolean,
+) {
+  const entry = loadSandboxRegistry().sandboxes[sandboxName];
+  if (!entry) {
+    throw new Error(`full E2E sandbox '${sandboxName}' is missing from the registry`);
+  }
+
+  const managedAuthority = readManagedWorkloadAuthority(entry);
+  if (managedAuthority) {
+    if (usedBuildKitPrebuild) {
+      throw new Error("managed-image cold onboarding must not use a local BuildKit prebuild");
+    }
+    return {
+      kind: managedAuthority.receipt.kind,
+      reference: managedAuthority.receipt.reference,
+      sourceCohort: managedAuthority.receipt.sourceCohort,
+      sourceRevision: managedAuthority.receipt.sourceRevision,
+    } as const;
+  }
+
+  if (entry.workload?.kind !== "legacy-dockerfile") {
+    throw new Error("full E2E cold onboarding must register a supported workload receipt");
+  }
+  if (!usedBuildKitPrebuild) {
+    throw new Error("legacy Dockerfile cold onboarding must use the local BuildKit prebuild");
+  }
+  return {
+    kind: entry.workload.kind,
+    reference: entry.workload.reference,
+  } as const;
+}

test/e2e/live/full-e2e.test.ts

@@ -45,6 +45,7 @@ import {
   fullE2eInferenceProbeEvidence,
   runFullE2eInferenceProbe,
 } from "./full-e2e-inference-probe.ts";
+import { readFullE2eColdWorkloadEvidence } from "./full-e2e-workload-evidence.ts";
 import { runOpenClawLaunchReadinessLeaseTurns } from "./launch-agent-turn.ts";
 import { bindApprovedPrBaseForBaseImageComparison } from "./pr-base-comparison.ts";
 
@@ -263,6 +264,7 @@ async function assertColdOnboardPerformance(input: {
   const maxSilenceMs = maximumOutputSilenceMs(traceWindow, input.outputEvents);
   const maxSilenceSecs = Math.ceil(maxSilenceMs / 1_000);
   const rootEndToInstallCompletionMs = input.installCompletedAtMs - traceWindow.finishedAtMs;
+  const workload = readFullE2eColdWorkloadEvidence(SANDBOX_NAME, usedBuildKitPrebuild);
 
   const firstTurnStartedAtMs = Date.now();
   const turn = await input.sandbox.execShell(
@@ -333,13 +335,13 @@ async function assertColdOnboardPerformance(input: {
     maxSilenceBudgetSecs: MAX_SILENCE_SECS,
     buildKitFallback,
     usedBuildKitPrebuild,
+    workload,
     classicBuildSteps,
     responseChars,
   });
 
   expect(plain, "expected literal wizard step [1/8] in installer output").toContain("[1/8]");
   expect(buildKitFallback, "expected no fallback from BuildKit to the gateway builder").toBe(false);
-  expect(usedBuildKitPrebuild, "expected the cold install to use BuildKit").toBe(true);
   expect(classicBuildSteps, "expected no classic per-instruction build steps").toBe(0);
   expect(
     maxSilenceSecs,

test/e2e/live/hermes-gpu-startup-proof.ts

@@ -51,9 +51,6 @@ export async function assertHermesGpuStartupProof({
   expect(installText).not.toContain("Reusing existing Docker-driver gateway");
   expect(installText).not.toContain("[reuse] Skipping gateway (running)");
   if (gpuRoute === "compatibility-only") {
-    expect(installText).toContain(
-      "Recreating OpenShell Docker sandbox container with NVIDIA GPU access",
-    );
     expect(installText).toContain("Docker container mode selected:");
     for (const fragment of HERMES_GPU_FALLBACK_DISCLOSURE_FRAGMENTS) {
       expect(installText).not.toContain(fragment);
@@ -65,20 +62,11 @@ export async function assertHermesGpuStartupProof({
     for (const fragment of HERMES_GPU_FALLBACK_DISCLOSURE_FRAGMENTS) {
       expect(installText).toContain(fragment);
     }
-    expect(installText).toContain(
-      "Recreating OpenShell Docker sandbox container with NVIDIA GPU access",
-    );
     expect(installText).toContain("Docker container mode selected:");
   } else {
     expect(installText).toContain(
       "Direct sandbox GPU enabled; allowing OpenShell GPU policy enrichment.",
     );
-    expect(installText).not.toContain(
-      "Recreating OpenShell Docker sandbox container with NVIDIA GPU access",
-    );
-    expect(installText).toContain(
-      "Recreating OpenShell Docker sandbox container with restart-safe startup",
-    );
     expect(installText).toContain(
       "Docker container mode selected: persistent sandbox startup command",
     );