tune: raise context budget 700->1400 tokens + max_tokens 1200->1800 for 600-800 word answers

app/domains/ayurveda.yaml

@@ -21,7 +21,7 @@ retrieval:
 
 llm:
   temperature: 0.3
-  max_tokens: 1200
+  max_tokens: 1800
 
 prompts:
   system: |

app/services/engine.py

@@ -690,7 +690,7 @@ async def stream_answer(
 
         # Build prompt (history is trimmed + deduplicated inside _build_prompt)
         trimmed_history = history[-(history_turns * 2):]
-        condensed_chunks, _ = condense_context(chunks, max_context_tokens=700)
+        condensed_chunks, _ = condense_context(chunks, max_context_tokens=1400)
         messages = _build_prompt(config, query, condensed_chunks, trimmed_history)
 
         # ── Step 4: Tier 2 — Stream buffer for OUT_OF_SYLLABUS detection ────────────