docs: park MI50 --jinja tool-calling as an experiment (cloud is the copilot path)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docs/PARKED_IDEAS.md

@@ -68,6 +68,19 @@ Dec-2025 plan (in her memory). The whole point of the project: can she become a
   cost loops, self-brick), not sci-fi. The isolation makes the *fun* version
   (emergence) safe to pursue. Build the box, then open the door.
 
+## 🛠️ Tool-calling on the MI50 (free local agency)
+Launch the MI50 llama.cpp server with `--jinja` so the `local-GPU` backend can
+do function-calling, then add `"mi50"` to `chat.TOOL_BACKENDS`. Would let the
+poker copilot + journaling tools run free/local instead of on cloud.
+- **Why parked:** not needed — cloud (gpt-4o) drives tools reliably and a full
+  poker session costs ~$0.50–1. A local 32B calls tools less reliably (wrong
+  tool / bad args / narrates instead) and is slower (round-trips × ~18s/turn),
+  which is exactly wrong for live at-the-table logging. Cloud is also easier to
+  debug tools against.
+- **Do it as:** a deliberate experiment to A/B the local model's tool-calling
+  (fits the "own stack" arc), not a dependency. Small + reversible: recreate the
+  CT202 container command with `--jinja`, keep it reboot-resilient.
+
 ## 🛠️ Deterministic poker tooling (RTO + cfr-core)
 Wire Lyra to Brian's own GTO/solver projects so ICM, equities, and ranges come
 from real computation, never LLM guesses.