epimetheus: fix double research message + Haiku query tuning

- Skip RESEARCH: tag when Haiku pre-pass already searched (no double-fire)
- Haiku told to use 2-3 word queries (was generating 6+ word queries that returned 0)
- Engagement filter dropped to 0 (niche crypto tweets have low engagement)
- systemd ProtectSystem paths fixed (root cause of ALL write failures)

Pentagon-Agent: Epimetheus <3D35839A-7722-4740-B93D-51157F7D5E70>

telegram/bot.py

@@ -288,7 +288,7 @@ async def handle_research(msg, query: str, user):
     await msg.chat.send_action("typing")
 
     logger.info("Research: searching X for '%s'", query)
-    tweets = await search_x(query, max_results=15, min_engagement=3)
+    tweets = await search_x(query, max_results=15, min_engagement=0)
     logger.info("Research: got %d tweets for '%s'", len(tweets), query)
     if not tweets:
         await msg.reply_text(f"No recent tweets found for '{query}'.")
@@ -490,10 +490,12 @@ async def handle_tagged(update: Update, context: ContextTypes.DEFAULT_TYPE):
             haiku_prompt = (
                 f"Does this Telegram message need a live X/Twitter search to answer well? "
                 f"Only say YES if the user is asking about recent sentiment, community takes, "
-                f"what people are saying, or emerging discussions that wouldn't be in a knowledge base.\n\n"
+                f"what people are saying, or emerging discussions.\n\n"
                 f"Message: {text}\n\n"
+                f"If YES, provide a SHORT search query (2-3 words max, like 'P2P.me' or 'MetaDAO buyback'). "
+                f"Twitter search works best with simple queries — too many words returns nothing.\n\n"
                 f"Respond with ONLY one of:\n"
-                f"YES: [search query]\n"
+                f"YES: [2-3 word query]\n"
                 f"NO"
             )
             haiku_result = await call_openrouter("anthropic/claude-haiku-4.5", haiku_prompt, max_tokens=50)
@@ -616,12 +618,14 @@ IMPORTANT: Two special tags you can append at the end of your response (after yo
             logger.info("Auto-learned [%s]: %s", category, correction[:80])
 
     # Auto-research (Ganymede: LLM-driven research trigger)
+    # Skip if Haiku pre-pass already searched (prevents double-fire + duplicate "No tweets found" messages)
     research_lines = re.findall(r'^RESEARCH:\s+(.+)$', response, re.MULTILINE)
     if research_lines:
         display_response = re.sub(r'\nRESEARCH:\s+.+$', '', display_response, flags=re.MULTILINE).rstrip()
-        for query in research_lines:
-            asyncio.get_event_loop().create_task(handle_research(msg, query.strip(), user))
-            logger.info("Auto-research triggered: %s", query[:80])
+        if not research_context:  # Only fire if Haiku didn't already search
+            for query in research_lines:
+                asyncio.get_event_loop().create_task(handle_research(msg, query.strip(), user))
+                logger.info("Auto-research triggered: %s", query[:80])
 
     # Post response (without LEARNING lines)
     await msg.reply_text(display_response)