fix(deploy): 启动脚本等端口真正释放，并确认服务可用后才报成功

上一次部署里旧 worker 没随 master 退出，一直占着 8123。新 master 启动、
打印 started、然后什么都不服务——日志每一行看起来都正常，站点却是挂的。

- stop.sh 等进程真正退出，超时后 -9
- start.sh 停完等端口释放，起来后 curl 健康检查通过才算成功，否则返回非零

Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>

deploy/start.sh

@@ -0,0 +1,47 @@
+#!/bin/sh
+# Start Garmin Health Lab. Safe to run repeatedly: an already-running
+# instance is stopped first. Intended for DSM Task Scheduler (boot-up).
+APP="$(cd "$(dirname "$0")" && pwd)"
+cd "$APP/backend" || exit 1
+
+GUNICORN="$APP/backend/.venv/bin/gunicorn"
+
+# Stop whatever is already running, then WAIT for the port to actually be
+# free. Killing only the pid in app.pid left orphaned workers holding :8123;
+# the new master then started, reported success, and served nothing — the
+# site was down while every log line looked normal.
+"$APP/stop.sh" >/dev/null 2>&1
+i=0
+while [ $i -lt 20 ]; do
+  pgrep -f "$GUNICORN" >/dev/null 2>&1 || break
+  sleep 1
+  i=$((i + 1))
+done
+pkill -9 -f "$GUNICORN" 2>/dev/null
+sleep 1
+
+mkdir -p "$APP/logs"
+# --timeout 300: an AI generation against the reasoning model can run for
+# minutes, and gunicorn kills a worker that looks stuck before then.
+nohup "$GUNICORN" \
+  --workers 2 --threads 4 --timeout 300 \
+  --bind 0.0.0.0:8123 \
+  --access-logfile "$APP/logs/access.log" \
+  --error-logfile "$APP/logs/error.log" \
+  wsgi:app > "$APP/logs/stdout.log" 2>&1 &
+
+echo $! > "$APP/app.pid"
+
+# Confirm it is actually serving rather than just running.
+i=0
+while [ $i -lt 25 ]; do
+  if curl -sf -m 2 -o /dev/null http://127.0.0.1:8123/api/health/status; then
+    echo "started pid $(cat "$APP/app.pid") on :8123"
+    exit 0
+  fi
+  sleep 1
+  i=$((i + 1))
+done
+
+echo "FAILED to serve on :8123 — see $APP/logs/error.log" >&2
+exit 1

deploy/stop.sh

@@ -0,0 +1,19 @@
+#!/bin/sh
+APP="$(cd "$(dirname "$0")" && pwd)"
+GUNICORN="$APP/backend/.venv/bin/gunicorn"
+
+if [ -f "$APP/app.pid" ]; then
+  kill "$(cat "$APP/app.pid")" 2>/dev/null && echo "stopped $(cat "$APP/app.pid")"
+  rm -f "$APP/app.pid"
+fi
+
+# The master's workers do not always go with it, and a surviving worker keeps
+# :8123 bound — which looks exactly like a healthy start that serves nothing.
+i=0
+while [ $i -lt 15 ]; do
+  pgrep -f "$GUNICORN" >/dev/null 2>&1 || exit 0
+  sleep 1
+  i=$((i + 1))
+done
+pkill -9 -f "$GUNICORN" 2>/dev/null
+echo "force-stopped leftover workers"