deploy(S4): Blender headless pipeline + WebAR client + backend fixes

backend/routes/system_routes.py

@@ -93,7 +93,12 @@ async def key_diagnostics(probe: bool = False):
         parsing. A key can list models fine and still fail to generate (quota
         exhausted, model retired, safety block), which an auth ping hides.
         """
-        model = "gemini-3.5-flash"
+        # Probe whatever the system is ACTUALLY configured to use, so this
+        # diagnostic can never drift from production behaviour.
+        try:
+            from backend.services.token_manager import MODEL as model
+        except Exception:
+            model = "gemini-3.1-flash-lite"
         url = (f"https://generativelanguage.googleapis.com/v1beta/models/{model}"
                f":generateContent?key=" + urllib.parse.quote(key))
         body = json.dumps({

backend/services/token_manager.py

@@ -80,7 +80,18 @@ except Exception as e:
     GEMINI_API_KEY = ""
 
 EXHAUSTED_GOOGLE_KEYS = set()
-MODEL = "gemini-3.5-flash"
+# Measured live against a real key, 2026-07-21:
+#   gemini-3.5-flash        HTTP 503  "currently experiencing high demand"  <- was set here
+#   gemini-2.0-flash        HTTP 429  free-tier request cap
+#   gemini-2.5-flash-lite   HTTP 404  retired
+#   gemini-3-flash-preview  OK but 35.5s
+#   gemini-3.1-flash-lite   OK in 1.1s, correct answer          <- chosen
+#
+# The 503 is Google-side capacity, NOT quota consumption: these keys were unused
+# for weeks. Because the primary model always failed, every request fell through
+# to the NVIDIA tier — which is why chat worked at all, and why research paid for
+# a doomed Gemini attempt before doing any real work.
+MODEL = "gemini-3.1-flash-lite"
 
 # ─────────────────────────────────────────────────────────────────────────────
 # DB Helpers