Deep inspection: document implementation gaps, fix critical bugs, add test coverage

[Copilot]

Repository claimed 5 complete benchmarks but delivered 1 complete + 2 stubs. No tests, broken imports, silent bugs. This PR audits implementation status, fixes critical issues, and establishes testing foundation.

Analysis & Documentation

Implementation Reality: ~30% complete vs README claims

• INSPECTION_REPORT.md - 10-section analysis mapping claimed features to actual code
• STATUS.md - Component-by-component completeness matrix
• TODO.md - Prioritized backlog with effort estimates
• SETUP.md - Installation and usage patterns

Key gaps identified:

• Benchmarks: 1/5 complete (Tool Maze ✓, Memory Drift stub, Recursive Planning stub, 2 missing)
• Telemetry: 1/5 metrics (GPU memory only)
• GPU features: FAISS only, no batch execution or vectorized rollouts

Bug Fixes

PlannerAgent infinite replanning

# Before: replanned when exhausted
def act(self, observation):
    if not self._plan:  # True both initially AND after exhaustion
        self._plan = list(self.planner(observation))
    ...

# After: one-time planning with fallback
def act(self, observation):
    if not self._has_planned:
        self._plan = list(self.planner(observation))
        self._has_planned = True
    if self._plan:
        return self._plan.pop(0)
    return self.fallback(observation)  # Now actually reached

Import resolution - Added __init__.py to all packages, sys.path setup in runners

Test Infrastructure

13 tests covering core components (all passing):

• Agents: ReactiveAgent, MemoryAgent, PlannerAgent behavior + fallback regression test
• Environments: Tool Maze success/failure/exhaustion, stub validation
• Memory: normalization, seeding, FAISS indexing/search

Package Structure

agents/__init__.py       # Export Agent, ActionResult, ReactiveAgent, MemoryAgent, PlannerAgent
environments/__init__.py # Export ToolMaze, MemoryDrift, RecursivePlanner
memory/__init__.py       # Export GpuFaissIndex, normalize_embeddings, seed_everything
telemetry/__init__.py    # Export query_vram, time_block
plots/__init__.py        # Export plot_metric
benchmarks/__init__.py   # Export run_tool_maze, main
tests/                   # 13 unit tests + runner

What Works Now

• ✅ All agent types functional (bugs fixed, tested)
• ✅ Tool Maze benchmark complete (tested)
• ✅ FAISS GPU/CPU fallback (tested)
• ✅ Import system without PYTHONPATH
• ✅ Test harness for safe iteration

What Remains

Per TODO.md priority order:

1. Memory Drift full implementation (sliding window, actual recall testing)
2. Recursive Planning full implementation (tree search, optimal solutions)
3. Deception Detection environment
4. Energy Budget tracking
5. Telemetry dashboard (tokens/sec, planning depth, memory growth, cost proxy)
6. GPU batch execution and vectorized rollouts

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.