fix(e2e): replace broken fault injection in onboard-repair and onboard-resume tests

test/e2e/test-onboard-repair.sh

@@ -132,14 +132,22 @@ pass "Exported NVIDIA_API_KEY for the repair run (host writes nothing to disk; O
 # Phase 2: Create interrupted resumable state
 # ══════════════════════════════════════════════════════════════════
 section "Phase 2: Create interrupted state"
-info "Running onboard with an invalid policy mode to create resumable state..."
+info "Running onboard with POLICY_MODE=custom but no POLICY_PRESETS to force a policy-step failure..."
 
+# Use NEMOCLAW_POLICY_MODE=custom without NEMOCLAW_POLICY_PRESETS — this is a
+# real validation path that exits 1 at the policy step after the sandbox is
+# already created, leaving resumable session state.
+#
+# Note: the previous approach (NEMOCLAW_POLICY_MODE=invalid) stopped working
+# after PR #2434 changed invalid modes from process.exit(1) to a graceful
+# fallback with console.warn(). See #2573 for details.
 FIRST_LOG="$(mktemp)"
 NEMOCLAW_NON_INTERACTIVE=1 \
   NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1 \
   NEMOCLAW_SANDBOX_NAME="$SANDBOX_NAME" \
   NEMOCLAW_RECREATE_SANDBOX=1 \
-  NEMOCLAW_POLICY_MODE=invalid \
+  NEMOCLAW_POLICY_MODE=custom \
+  NEMOCLAW_POLICY_PRESETS="" \
   node "$REPO/bin/nemoclaw.js" onboard --non-interactive >"$FIRST_LOG" 2>&1
 first_exit=$?
 first_output="$(cat "$FIRST_LOG")"
@@ -159,7 +167,7 @@ else
   fail "Onboard session file missing after interrupted run"
 fi
 
-if echo "$first_output" | grep -q "Unsupported NEMOCLAW_POLICY_MODE: invalid"; then
+if echo "$first_output" | grep -q "NEMOCLAW_POLICY_PRESETS is required when NEMOCLAW_POLICY_MODE=custom"; then
   pass "First run failed at policy setup as intended"
 else
   fail "First run did not fail at the expected policy step"
@@ -223,7 +231,8 @@ else
   fail "Repair resume did not report missing sandbox recreation"
 fi
 
-if echo "$repair_output" | grep -q "\[5/7\] Creating sandbox"; then
+# The step numbering is [6/8] in the current onboard flow.
+if echo "$repair_output" | grep -q "Creating sandbox"; then
   pass "Repair resume recreated sandbox"
 else
   fail "Repair resume did not rerun sandbox creation"
@@ -239,6 +248,22 @@ fi
 # Phase 4: Reject conflicting sandbox
 # ══════════════════════════════════════════════════════════════════
 section "Phase 4: Reject conflicting sandbox"
+
+# Phase 3 completed the session (resumable=false). Re-create interrupted state
+# so the conflict detection path is exercised (it runs before the "no resumable
+# session" early-exit).
+info "Re-creating interrupted state for conflict testing..."
+REINJECT_LOG="$(mktemp)"
+NEMOCLAW_NON_INTERACTIVE=1 \
+  NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1 \
+  NEMOCLAW_SANDBOX_NAME="$SANDBOX_NAME" \
+  NEMOCLAW_RECREATE_SANDBOX=1 \
+  NEMOCLAW_POLICY_MODE=custom \
+  NEMOCLAW_POLICY_PRESETS="" \
+  node "$REPO/bin/nemoclaw.js" onboard --non-interactive >"$REINJECT_LOG" 2>&1 || true
+rm -f "$REINJECT_LOG"
+pass "Re-created interrupted session for conflict tests"
+
 info "Attempting resume with a different sandbox name..."
 
 SANDBOX_CONFLICT_LOG="$(mktemp)"

test/e2e/test-onboard-resume.sh

@@ -142,14 +142,22 @@ pass "Exported NVIDIA_API_KEY for the resume run (host writes nothing to disk; O
 # Phase 2: First onboard (forced failure after sandbox creation)
 # ══════════════════════════════════════════════════════════════════
 section "Phase 2: First onboard (interrupted)"
-info "Running onboard with an invalid policy mode to create resumable state..."
+info "Running onboard with POLICY_MODE=custom but no POLICY_PRESETS to force a policy-step failure..."
 
+# Use NEMOCLAW_POLICY_MODE=custom without NEMOCLAW_POLICY_PRESETS — this is a
+# real validation path that exits 1 at the policy step after the sandbox is
+# already created, leaving resumable session state.
+#
+# Note: the previous approach (NEMOCLAW_POLICY_MODE=invalid) stopped working
+# after PR #2434 changed invalid modes from process.exit(1) to a graceful
+# fallback with console.warn(). See #2573 for details.
 FIRST_LOG="$(mktemp)"
 NEMOCLAW_NON_INTERACTIVE=1 \
   NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1 \
   NEMOCLAW_SANDBOX_NAME="$SANDBOX_NAME" \
   NEMOCLAW_RECREATE_SANDBOX=1 \
-  NEMOCLAW_POLICY_MODE=invalid \
+  NEMOCLAW_POLICY_MODE=custom \
+  NEMOCLAW_POLICY_PRESETS="" \
   node "$REPO/bin/nemoclaw.js" onboard --non-interactive >"$FIRST_LOG" 2>&1
 first_exit=$?
 first_output="$(cat "$FIRST_LOG")"
@@ -169,7 +177,7 @@ else
   fail "Sandbox creation not confirmed in first run output"
 fi
 
-if echo "$first_output" | grep -q "Unsupported NEMOCLAW_POLICY_MODE: invalid"; then
+if echo "$first_output" | grep -q "NEMOCLAW_POLICY_PRESETS is required when NEMOCLAW_POLICY_MODE=custom"; then
   pass "First run failed at policy setup as intended"
 else
   fail "First run did not fail at the expected policy step"
@@ -261,10 +269,15 @@ else
   pass "Resume did not rerun sandbox creation"
 fi
 
+# The first onboard completed through openclaw (step 7) before failing at
+# policies (step 8). Inference was already configured during that run, so
+# the resume path detects it is ready (isInferenceRouteReady) and skips it.
 if echo "$resume_output" | grep -q "\[4/7\] Setting up inference provider"; then
-  pass "Resume continued with inference setup"
+  pass "Resume re-ran inference setup"
+elif echo "$resume_output" | grep -q "\[resume\] Skipping inference\|\[reuse\] Skipping inference"; then
+  pass "Resume skipped inference (already configured)"
 else
-  fail "Resume did not continue with inference setup"
+  fail "Resume neither ran nor skipped inference setup"
 fi
 
 if run_nemoclaw "$SANDBOX_NAME" status >/dev/null 2>&1; then