safety: cap threads at CPU_THREADS_MAX (default 2) to prevent oversubscription if detection over-reports

app.py

@@ -57,8 +57,12 @@ def _effective_cpus():
 
 
 _CPU = _effective_cpus()
-CPU_THREADS = int(os.getenv("CPU_THREADS", str(_CPU)))
-CPU_THREADS_BATCH = int(os.getenv("CPU_THREADS_BATCH", str(_CPU)))
+# Hard ceiling on threads: even if detection over-reports (e.g. host cores leak
+# through), never oversubscribe the ~2-vCPU tier. Raise CPU_THREADS_MAX if you
+# move to a bigger CPU tier.
+CPU_THREADS_MAX = int(os.getenv("CPU_THREADS_MAX", "2"))
+CPU_THREADS = min(int(os.getenv("CPU_THREADS", str(_CPU))), CPU_THREADS_MAX)
+CPU_THREADS_BATCH = min(int(os.getenv("CPU_THREADS_BATCH", str(_CPU))), CPU_THREADS_MAX)
 CONTEXT_SIZE = int(os.getenv("CONTEXT_SIZE", "8192"))   # incl. image tokens; per-slot = ctx/parallel
 PARALLEL = int(os.getenv("PARALLEL", "2"))
 BATCH_SIZE = int(os.getenv("BATCH_SIZE", "512"))