Count a dead letter by when it was kept, so one kept late still counts (issue 440)

The give-up time is not newer for every letter kept: a notice that could
not be kept is offered again a minute later, so a later give-up can be kept
first and the one after it read as not fresh. The stored time rises with the
stream's sequence. A consumer whose letters vanish between the two reads is
left out of the look instead of counted as a look, and the skip of a
token-less max-deliveries advisory now has a test of its own.

cmd/mesh-controller/signals.go

@@ -576,9 +576,11 @@ func watchDeadLetters(f *signalFacts) []conditions.Observation {
 			messages, them = fmt.Sprintf("%d messages", n), "them"
 		}
 		who := consumerWho(stream, consumer)
-		// DEAD_LETTERS is a record of what was given up on: the condition counts the messages and says when
-		// the newest was given up, never how often the controller looked (novox/hq issues 402 and 440).
-		// Without that time it counts its looks, as a source of the present does.
+		// DEAD_LETTERS is a record of what was given up on: the condition says when the newest held message
+		// was kept, and its count is a running total of the messages given up on while it is open, never
+		// how often the controller looked (novox/hq issues 402 and 440). It is at least the number held now,
+		// and does not go down when one is delivered again or dropped. Without that time it counts its
+		// looks, as a source of the present does.
 		happened, times := f.deadLettersNewest[key], 0
 		if !happened.IsZero() {
 			times = n