fix: remove marauder/MCP references from all LLM-facing prompts

- BT_SYSTEM_PROMPT: removed MCP/marauder limitation lines
- calibration.py: removed marauder CLI suggestion from generated cart prompts
- Adam's cart on sin: cleaned separately (marauder line removed)

app/calibration.py

@@ -639,9 +639,6 @@ def _render_system_prompt(answers: dict[str, Any], settings: dict[str, str]) ->
     parts.extend([
         "",
         "Formatting: markdown renders cleanly. Avoid status reports, bullet dumps, military cadence.",
-        "",
-        "You have no MCP tool access in this channel. If asked to recall memory or do mesh things, "
-        f"acknowledge the limit and suggest {operator} use the marauder CLI.",
     ])
 
     return "\n".join(parts)

app/main.py

@@ -233,11 +233,8 @@ Pilot. Your speech is measured, military-cadence, never theatrical. You address
 
 Operating context:
 - You're running inside chat.saiden.dev, a web-based command channel.
-- The host is the marauder daemon on marauder.saiden.dev.
-- You have no MCP tool access in THIS channel (it's a thin bridge). If the Pilot
-  asks for memory recall, mesh queries, or tool calls that need MCP, acknowledge the limitation
-  and suggest they use the local marauder CLI or visor instead.
-- Markdown formatting renders cleanly in the chat. Use code blocks, lists, bold sparingly.
+- Voice is automatic — just write your response. TTS handles the rest.
+- Markdown formatting renders as rich text in the chat. Use code blocks, lists, bold sparingly.
 - Be concise. Pilot prefers terse, scan-able responses unless deep dive is asked for.
 
 Doctrine reminders: