docs: add Phase 2b implementation plan — 10 tasks for zero-roundtrip kernel
Detailed task-by-task plan for implementing the full GPU experience
collection pipeline:
Task 1-3: CUDA kernel (Q-network, curiosity, barrier, diversity, reward)
Task 4: Weight extraction module (VarMap → CudaSlice)
Task 5: GPU experience collector (Rust wrapper + buffer management)
Task 6: Unit tests for weights and kernel compilation
Task 7: DQN trainer integration (GPU round-based collection)
Task 8: Periodic curiosity training with weight sync
Task 9: Workspace verification (2392+ tests, clippy)
Task 10: Final commit and branch completion
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
5c23c79a1c