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.

README.md

@@ -105,10 +105,18 @@ reported while the commits stay put: nothing in Gitea's payload separates it fro
 a merge check still in flight, and the baseline line is the only notice of it.
 That is a narrower gap than it looks, because `base.sha` is the base branch's
 tip — it moves for every open PR whenever anything merges to the base branch, so
-the commits rarely stay put for long. A move like that arms the rule, and the
-move itself is silent: the window restarts from it, so what is eventually
-reported is a conflict still standing two minutes after that fresh merge
-computation began, never the run of falses that preceded it.
+the commits rarely stay put for long. The first such move arms the rule, and is
+itself silent: the window then runs from that arming poll rather than from the
+falses that predate it, so what is eventually reported is a conflict that
+outlasted the merge computation the move started.
+
+Arming is a one-way latch. Moves after that one — and every move seen by a watch
+that was already armed at the baseline — neither re-arm nor restart the window,
+because a restart on every `base.sha` move could never complete on a busy base
+branch. That is the cost side of the same trade: on a busy base branch a
+conflict can be confirmed while the newest merge recompute is less than two
+minutes old. The window guarantees that the run of falses outlasted *a* merge
+computation this watch saw start, not that it outlasted the most recent one.
 
 ```bash
 # Watch until something meaningful happens (default interval 60s)

cmd/watchpr/main.go

@@ -8,8 +8,12 @@
 // so. A conflict introduced by the push immediately before the watch started is
 // not reported while the commits stay put, because nothing in Gitea's payload
 // separates it from a merge check still in flight; the baseline line is the only
-// notice of it. A later move of the head or base arms the rule without alerting,
-// and restarts the two-minute window from itself.
+// notice of it. The first later move of the head or base arms the rule without
+// alerting, and the window then runs from that arm rather than from the falses
+// that predate it. Moves after that neither re-arm nor restart the window -- a
+// restart on every base move could never complete on a busy base branch -- so a
+// conflict can be confirmed while the newest merge recompute is less than two
+// minutes old.
 //
 //	watchpr owner/repo#12 owner/repo:15
 //	watchpr --once --json owner/repo#12
@@ -108,9 +112,11 @@ func clientFor(jsonMode bool) *agent.GiteaClient {
 	return agent.NewGiteaClient(token)
 }
 
-// warnRecord is the --json form of a warning. Under --json stderr carries
-// nothing but NDJSON records, so a caller parsing it line by line never has to
-// guess which shape a line is.
+// warnRecord is the --json form of a warning. Under --json every notice watchpr
+// itself writes to stderr -- warnings and the baseline -- is an NDJSON record,
+// so a caller parsing them line by line never has to guess which shape a line
+// is. A terminal failure is the one exception: SilenceErrors stays off, so
+// cobra prints it as a plain "Error: ..." line and the exit status is non-zero.
 type warnRecord struct {
 	Warning string `json:"warning"`
 }
@@ -255,7 +261,7 @@ func suppressedAtBaseline(st agent.PRState) string {
 	case agent.MergeNo:
 		conds = append(conds, "already non-mergeable (a real conflict, or Gitea still recomputing)")
 	case agent.MergeUnknown:
-		conds = append(conds, "mergeability unknown (the conflict rule is disarmed until Gitea answers or the commits move)")
+		conds = append(conds, "mergeability unknown (the conflict rule is disarmed until Gitea reports mergeable or the commits move)")
 	}
 	if st.CIStatus == "failure" || st.CIStatus == "error" {
 		conds = append(conds, "CI already "+st.CIStatus)

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",