fix: Resolve remaining test and documentation issues

- Update verify_capabilities.py to unpack 3-tuple from deliberate().
- Remove unused matplotlib import from test_architecture_stability.py.
- Update ROADMAP.md with C++ inference optimization task.

Author: Pomilon

docs/ROADMAP.md

@@ -76,6 +76,9 @@ This phase focuses on scaling the model and performing rigorous, comparative ben
             *   A baseline Mamba model of the same size (MCTS disabled).
             *   A standard Transformer model of a similar size.
         *   **[TODO]** Measure and compare latency, throughput, and computational cost for generation.
+    *   **3. Inference Optimization (C++/Rust):**
+        *   **[TODO]** Port the MCTS Planner logic to C++ (LibTorch) or Rust to bypass the Python GIL.
+        *   **[TODO]** Implement true parallel execution where "System 2" does not block "System 1" generation.
 
 *   **Exit Criteria:**
     *   A large-scale CRSM model is successfully trained.

tests/test_architecture_stability.py

@@ -4,7 +4,6 @@
 import numpy as np
 import sys
 import os
-import matplotlib.pyplot as plt
 
 # Add project root to path
 sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))