Convict on a released assignment instead of skipping

The pre-conviction recheck treated a missing assignment like a changed
one and skipped the quarantine. A stop or delete landing between the
metadata snapshot and the recheck releases the assignment but cannot
un-wedge the VF, and the archive-before-persist ordering already ties a
marker read under this epoch's snapshot to this epoch's VF, so the
conviction is valid. Skip only when a different assignment exists.

Author: yummybomb

lib/instances/vgpu_sentinel.go

@@ -231,19 +231,20 @@ func (c *VGPUSentinelController) convict(ctx context.Context, target vgpuSentine
 	// The target came from a metadata snapshot taken before the log read: a
 	// stop/start racing the scan could have assigned a different VF by now,
 	// and convicting the snapshot's VF would quarantine a healthy device.
-	// Reload the assignment and convict only if it is unchanged. Start
+	// Reload the assignment and skip only if a different one exists. Start
 	// archives the previous boot's log before persisting a new assignment,
-	// so if the metadata still shows this epoch after the marker was read,
-	// the marker provably belongs to this epoch's VF. On any other outcome
-	// the tail stays open; the next scan sees the current assignment and
-	// rescans its log from the top.
+	// so a marker read under this epoch's snapshot provably belongs to this
+	// epoch's VF — including when the instance was stopped or deleted after
+	// the read, which releases the assignment but cannot un-wedge the VF.
+	// On a skip or error the tail stays open; the next scan sees the current
+	// assignment and rescans its log from the top.
 	current, ok, err := c.store.getVGPUSentinelTarget(ctx, target.instanceID)
 	if err != nil {
 		c.log.WarnContext(ctx, "vGPU sentinel could not confirm assignment before conviction",
 			"vf", target.vfAddress, "instance_id", target.instanceID, "error", err)
 		return false
 	}
-	if !ok || current.vfAddress != target.vfAddress || current.assignedAt != target.assignedAt {
+	if ok && (current.vfAddress != target.vfAddress || current.assignedAt != target.assignedAt) {
 		c.log.InfoContext(ctx, "vGPU sentinel skipping conviction: assignment changed during scan",
 			"vf", target.vfAddress, "instance_id", target.instanceID)
 		return false

lib/instances/vgpu_sentinel_test.go

@@ -268,11 +268,13 @@ func TestVGPUSentinelControllerSkipsConvictionOnChangedAssignment(t *testing.T)
 	assert.Empty(t, *quarantined)
 	assert.False(t, c.tails["instance-1"].done)
 
-	// A deleted instance can no longer vouch for the marker either.
+	// A stop or delete after the marker was read releases the assignment but
+	// cannot un-wedge the VF: the marker still convicts the scanned epoch's VF.
 	store.targets = nil
 	c.tails["instance-1"] = &vgpuSentinelTail{vfAddress: stale.vfAddress, assignedAt: stale.assignedAt}
 	c.scanTarget(context.Background(), stale)
-	assert.Empty(t, *quarantined)
+	require.Len(t, *quarantined, 1)
+	assert.Equal(t, "0000:e3:00.4", (*quarantined)[0].VFAddress)
 }
 
 func TestListVGPUSentinelTargetsSkipsUnstattableMetadata(t *testing.T) {