**feat: refactor agent loop to state machine with dynamic trajectory checks**

- Replace hard-coded `MAX_REPL_STEPS` death-limit with a dynamic Checkpoint Interval.
- Introduce 4-phase state machine: `PLANNING`, `EXECUTING`, `TRAJECTORY_CHECK`, and `PIVOTING`.
- Add Trajectory Checks allowing the agent to self-assess progress against defined success criteria to extend its execution budget.
- Implement Hard (wipe strategy/kernel) and Soft (retry step) Pivoting logic.
- Establish persistent notebook-style kernel (flushed only on Hard Pivots).
- Upgrade sandbox safety with strict import whitelisting, dunder blocking, and taint tracking.

**bugfix:

- logging_config.py: Add markdown_it to suppressed logger list
- utils.py: Rewrite save_agent_trace for delta-based logging
- edge_rlm.py: Add last_trace_msg_count tracking and pass deltas at all 3 call sites

.env.example

@@ -0,0 +1,11 @@
+# LLM API endpoints
+AGENT_API=http://localhost:8080/v1
+REPL_API=http://localhost:8090/v1
+
+# File paths
+CONTEXT_FILE=context.txt
+TASK_FILE=task.txt
+
+# Agent behaviour
+MAX_REPL_STEPS=20
+MAX_VIRTUAL_CONTEXT_RATIO=0.85