agent: bump default MEMORY_TOKEN_THRESHOLD 16k -> 80k

QUEST-35B serves with max_model_len=160000, so 16k was triggering the
condenser way earlier than needed and hurting throughput. 80k gives the
agent more headroom to accumulate tool responses before paying the
summarizer LLM round-trip; condenser still kicks in well before the
context window cap.

app.py

@@ -1564,7 +1564,7 @@ MEMORY_TOKEN_THRESHOLD = int(
     os.getenv("MEMORY_THRESHOLD")
     or os.getenv("MEMORY_CONTEXT_THRESHOLD")
     or os.getenv("MEMORY_TOKEN_THRESHOLD")
-    or "16000"
+    or "80000"
 )
 
 # Azure OpenAI support — mirrors inference/tool_visit.py logic. When