watchpr: describe arming as a one-way latch, not a window restart

The docs claimed every later head/base move restarts the conflict window.
arm() returns early once armed, so only the first move sets armedAt and
every move after it is a no-op. State the real trade-off instead: repeated
moves do not extend the debounce, so a conflict can be confirmed while the
newest recompute is younger than the window.

Also narrow the --json stderr claim to the notices watchpr writes itself
(cobra's terminal Error: line is plain text), rename the test to what it
covers, and stop the unknown-mergeability baseline implying a false answer
arms the rule.

cmd/watchpr/main_test.go

@@ -257,10 +257,12 @@ func TestBaselineIsEmittedInJSONMode(t *testing.T) {
 	}
 }
 
-// Under --json stderr is the baseline and warning channel, so it has to be one
-// shape: a caller parsing it line by line must never meet a bare `warning:`
-// line between two NDJSON records.
-func TestJSONModeStderrIsAllRecords(t *testing.T) {
+// Under --json stderr is the baseline and warning channel, so every notice
+// watchpr writes there has to be one shape: a caller parsing it line by line
+// must never meet a bare `warning:` line between two NDJSON records. Cobra's
+// terminal `Error: ...` line is not covered here -- SilenceErrors stays off, so
+// it is plain text on stderr alongside a non-zero exit.
+func TestJSONModeWarningsAndBaselineAreRecords(t *testing.T) {
 	st := agent.PRState{
 		Ref:       agent.PRRef{Owner: "unkin", Repo: "repo", Number: 7},
 		State:     "open",