fix(panel): bed om en planlagt genstart, ikke en der dør med afsenderen

Knappen tog produktionen ned og lod den ligge. Panelets almindelige restart
udfører arbejdet mens forespørgslen står åben: den stopper containeren — hvilket
slår denne proces ihjel — og panelet, der ser afsenderen forsvinde, når aldrig
til den halvdel, der starter igen. Fire minutters nedetid den 18. august, og det
kan ikke lappes herfra: enhver knap, der synkront beder om sin egen død, har
samme form.

safe-restart planlægges i stedet. Panelet svarer "Restart scheduled" og gør
arbejdet bagefter for egen regning, så intet afhænger af, at denne proces stadig
er i live til at høre svaret.

Prøven hang bagefter, og det var prøvens skyld: dens falske panel læste aldrig
forespørgslens krop, og der er en nu. En server, der ikke læser hvad den fik,
opdager aldrig at afsenderen har lagt på — så httptest ventede på en handler, der
aldrig vendte tilbage, og hele binæren løb tør for tid i stedet for den ene
forespørgsel.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

internal/httpapi/panel_handlers.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"fmt"
 	"net/http"
+	"strings"
 	"time"
 )
 
@@ -72,13 +73,26 @@ func (s *Server) handleRestartFromPanel(w http.ResponseWriter, r *http.Request)
 	ctx, cancel := context.WithTimeout(context.WithoutCancel(r.Context()), 10*time.Second)
 	defer cancel()
 
-	url := fmt.Sprintf("%s/api/servers/%s/restart", s.cfg.PanelURL, s.cfg.PanelServerID)
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, nil)
+	// safe-restart rather than restart, and the difference is not politeness.
+	//
+	// The plain endpoint does its work while the request is open: it stops the
+	// container — which kills this process — and the panel, seeing the caller
+	// vanish, never gets to the starting half. That took production down and left
+	// it down for four minutes on 18 August, and it cannot be fixed from this side:
+	// any button that synchronously asks for its own death has the same shape.
+	//
+	// safe-restart is scheduled. The panel answers "Restart scheduled" and does the
+	// work on its own afterwards, so nothing depends on this process still being
+	// alive to hear the reply.
+	url := fmt.Sprintf("%s/api/servers/%s/safe-restart", s.cfg.PanelURL, s.cfg.PanelServerID)
+	body := strings.NewReader(`{"backup_first":false,"target_id":""}`)
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, body)
 	if err != nil {
 		s.internal(w, r, "build panel request", err)
 		return
 	}
 	req.Header.Set("Authorization", "Bearer "+s.cfg.PanelToken)
+	req.Header.Set("Content-Type", "application/json")
 
 	resp, err := (&http.Client{Timeout: 10 * time.Second}).Do(req)
 	if err != nil {

internal/httpapi/panel_test.go

@@ -1,6 +1,7 @@
 package httpapi
 
 import (
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -15,6 +16,11 @@ import (
 // request has to end either way, or whatever sits in front answers for it.
 func TestAPanelThatNeverAnswersStillEndsTheRequest(t *testing.T) {
 	panel := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// Drained first. A real server reads what it was sent, and one that does not
+		// never learns the caller has hung up — httptest.Server.Close then waits for
+		// a handler that will never return, and the whole binary times out instead
+		// of the one request.
+		_, _ = io.Copy(io.Discard, r.Body)
 		<-r.Context().Done()
 	}))
 	defer panel.Close()
@@ -56,6 +62,7 @@ func TestAPanelThatNeverAnswersStillEndsTheRequest(t *testing.T) {
 // And a panel that says no must say so in this server's own words, not in a proxy's.
 func TestAPanelThatRefusesSaysSoAsJSON(t *testing.T) {
 	panel := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = io.Copy(io.Discard, r.Body)
 		w.WriteHeader(http.StatusForbidden)
 	}))
 	defer panel.Close()