cpustress+orchestrator: serial CPU/RAM passes + silent-skip guard

Orion's run (log 20:49 → 20:54) shipped GREEN while silently skipping
CPUStress. Two compounding bugs:

1. CPUStress ran --cpu N AND --vm N --vm-bytes 90% concurrently.
   On a 4-core 8 GiB N95, that's 360% RAM overcommit; the OOM-killer
   fired, usually on the agent itself. Replaced with two sequential
   passes — CPU (all methods, --verify) for 3 min, then RAM (--vm 1,
   --vm-bytes capped to MemAvailable − 1.5 GiB, floor 256 MiB, --verify)
   for 3 min. Each pass now also asserts elapsed ≥ target − 2s so a
   premature clean exit counts as failure instead of a silent pass.

2. On systemd-restart after the OOM, the agent hardcoded nextStage :=
   "Inventory" and re-ran it. The orchestrator's /result handler
   advances run state via TriggerStageCompleted against the *current*
   RunState, not against body.Stage — so an Inventory result posted
   while the run was in StateCPUStress silently advanced CPUStress →
   Storage and marked CPUStress passed without it ever running.

Two-layer defense for #2:
- agent-side: /claim response now carries current_state; agent resumes
  at the matching stage on a re-claim (happy path).
- server-side: new TriggerStageMismatch + StageNameForState helper
  backstop. If body.Stage doesn't match the run's current stage, /result
  parks the run in FailedHolding with failed_stage labeled
  "<got> (expected <expected>)" and returns 409.

Other stages audited for similar unbounded concurrency — none found;
only CPUStress was unsafe.

Tests:
- cpustress_test.go — parseMemAvailable parses real meminfo, errors on
  missing/malformed; cap calc hits floor on tiny boxes, uses 1.5 GiB
  headroom on normal/huge boxes.
- statemachine_test.go — TriggerStageMismatch lands at FailedHolding
  from every stage state and is rejected from pre-stage/terminal
  states; StageNameForState round-trips the stageStates map.
- agent_handlers_test.go — TestResult_RejectsMismatchedStage proves
  the Orion scenario now 409s + FailedHolding; TestResult_AcceptsMatchingStage
  proves the guard doesn't break the happy path.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

internal/api/agent_handlers.go

@@ -180,6 +180,15 @@ func (a *Agent) Claim(w http.ResponseWriter, r *http.Request) {
 		}
 	}
 
+	// Re-fetch run state: the Transition above may have advanced us from
+	// Booting → InventoryCheck, and we want to hand that fresh state to
+	// the agent so a re-claim after a crash resumes at the stored state
+	// instead of silently replaying Inventory.
+	currentState := run.State
+	if fresh, err := a.Runs.Get(r.Context(), runID); err == nil && fresh != nil {
+		currentState = fresh.State
+	}
+
 	log.Printf("agent claimed: run=%d agent_ip=%s", runID, agentIP)
 	if a.Logs != nil {
 		if w, err := a.Logs.WriterFor(runID); err == nil {
@@ -213,6 +222,7 @@ func (a *Agent) Claim(w http.ResponseWriter, r *http.Request) {
 		"expected_disks":  expectedDisks,
 		"iperf_port":      iperfPort,
 		"non_destructive": run.NonDestructive,
+		"current_state":   string(currentState),
 	})
 }
 
@@ -411,6 +421,42 @@ func (a *Agent) Result(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// Silent-skip guard. Orchestrator advances the run state via
+	// TriggerStageCompleted against the *current* state, not against
+	// body.Stage — so an Inventory result posted while the run is in
+	// StateCPUStress would silently advance CPUStress → Storage and mark
+	// CPUStress as passed without it ever running. That's exactly what
+	// happened on Orion when the agent OOM-crashed mid-CPUStress,
+	// systemd restarted it, and the restarted agent (which hardcoded
+	// "Inventory" as its first stage) re-ran Inventory and reported it.
+	// Guard: if body.Stage doesn't match the stage the run is currently
+	// in, park the run in FailedHolding so the operator can investigate
+	// rather than trusting the claim and cascading silent passes.
+	expectedStage := orchestrator.StageNameForState(run.State)
+	if expectedStage != "" && body.Stage != expectedStage {
+		failedLabel := fmt.Sprintf("%s (expected %s)", body.Stage, expectedStage)
+		if err := a.Runs.SetFailedStage(r.Context(), runID, failedLabel); err != nil {
+			log.Printf("result: set failed stage on mismatch run %d: %v", runID, err)
+		}
+		if _, err := a.Runner.Transition(r.Context(), runID, orchestrator.TriggerStageMismatch); err != nil {
+			log.Printf("result: stage-mismatch transition run %d: %v", runID, err)
+		}
+		hostName := a.hostNameFor(r.Context(), run.HostID)
+		a.dispatchEvent(notify.Event{
+			Kind:     notify.KindStageFailed,
+			Severity: notify.SeverityCritical,
+			RunID:    runID,
+			HostName: hostName,
+			Title:    fmt.Sprintf("[vetting] %s stage mismatch: %s", hostName, body.Stage),
+			Body: fmt.Sprintf("Run %d reported stage %s while orchestrator expected %s — parked in FailedHolding to prevent silent skip.",
+				runID, body.Stage, expectedStage),
+			URL: a.runLinkURL(runID),
+		})
+		log.Printf("result: stage mismatch run=%d got=%s expected=%s — parked", runID, body.Stage, expectedStage)
+		http.Error(w, "stage mismatch: got "+body.Stage+", expected "+expectedStage, http.StatusConflict)
+		return
+	}
+
 	stageState := model.StagePassed
 	if !body.Passed {
 		stageState = model.StageFailed