fix(validator): tighten nvidia-smi driver version parsing

Require the version on the same banner line (horizontal whitespace only),
wrap log-verification errors with the node name, and add a numeric-order
floor case so lexical string compare cannot silently pass the tests.

Signed-off-by: Rohit Rajani <rorajani@nvidia.com>

Author: rorajani

validators/deployment/nvidia_smi.go

@@ -70,12 +70,14 @@ const (
 // nvidiaSMIDriverVersionRE extracts the host driver / KMD version from an
 // nvidia-smi banner. Matches both legacy ("Driver Version:") and renamed
 // ("KMD Version:") fields, case-insensitively, including the table-row layout
-// where fields sit on one pipe-delimited line (issue #1667). Caps at three
-// numeric components — NVIDIA driver versions are Major.Minor.Patch. Truncation
-// of a longer version (e.g. "580.95.05.1") is rejected in
-// parseNvidiaSMIDriverVersion: Go's RE2 engine has no negative lookahead.
+// where fields sit on one pipe-delimited line (issue #1667). After the colon,
+// only horizontal whitespace is allowed so a version on the next line is not
+// treated as the field's value. Caps at three numeric components — NVIDIA
+// driver versions are Major.Minor.Patch. Truncation of a longer version
+// (e.g. "580.95.05.1") is rejected in parseNvidiaSMIDriverVersion: Go's RE2
+// engine has no negative lookahead.
 var nvidiaSMIDriverVersionRE = regexp.MustCompile(
-	`(?i)(?:driver|kmd)\s+version:\s*([0-9]+(?:\.[0-9]+){0,2})`)
+	`(?i)(?:driver|kmd)\s+version:[ \t]*([0-9]+(?:\.[0-9]+){0,2})`)
 
 // gpuNodeCoverage partitions check-nvidia-smi's discovered GPU nodes into the
 // schedulable cohort actually validated and the cordoned cohort skipped. It
@@ -367,7 +369,8 @@ func verifySingleGPUNode(ctx *validators.Context, nodeName string) error {
 	}
 
 	if err := verifyNvidiaSMILogs(podLogs, createdPod); err != nil {
-		return err
+		return errors.Wrap(errors.ErrCodeInternal,
+			fmt.Sprintf("nvidia-smi log verification failed on node %s", nodeName), err)
 	}
 	return enforceGPUDriverVersionFloor(ctx, podLogs, nodeName)
 }

validators/deployment/nvidia_smi_test.go

@@ -236,6 +236,11 @@ func TestParseNvidiaSMIDriverVersion(t *testing.T) {
 			logs:    "Driver Version: 580.95.05-rc1\n",
 			wantErr: true,
 		},
+		{
+			name:    "rejects version on the next line after the field",
+			logs:    "Driver Version:\n580.95.05\n",
+			wantErr: true,
+		},
 		{
 			name: "accepts table pipe immediately after version",
 			logs: "| Driver Version: 580.95.05| CUDA Version: 12.8 |\n",
@@ -293,6 +298,13 @@ func TestEnforceGPUDriverVersionFloor(t *testing.T) {
 			constraint: ">= 580.95.05",
 			logs:       goodLogs,
 		},
+		{
+			// Lexical string compare would fail here ('.100' < '.99'); the
+			// constraint evaluator must compare components numerically.
+			name:       "numeric order beats lexical order",
+			constraint: ">= 580.99.99",
+			logs:       "NVIDIA-SMI\nDriver Version: 580.100.0\nCUDA Version: 12.8\n" + gpuCheckSuccessMsg,
+		},
 		{
 			name:       "below floor fails",
 			constraint: ">= 580.95.05",