Merge commit 'd803bfe2b1fe7f5e219e50ac20d6801a0a58ac75' as 'vendor/ruvector'

vendor/ruvector/crates/ruQu/docs/RESEARCH_DISCOVERIES.md

@@ -0,0 +1,210 @@
+# Research Discoveries for ruQu Enhancement
+
+*Compiled: January 2026*
+
+This document captures state-of-the-art research findings that can inform further improvements to ruQu's coherence gate architecture.
+
+---
+
+## 1. Real-Time Decoding at Scale
+
+### DECONET System (April 2025)
+**Source**: [arXiv:2504.11805](https://arxiv.org/abs/2504.11805)
+
+DECONET is a first-of-its-kind decoding system that scales to **thousands of logical qubits** with lattice surgery support. Key innovations:
+
+- **Network-integrated hybrid tree-grid structure**: O(log(l)) latency increase as system grows
+- **Resource scaling**: O(l × log(l)) compute, O(l) I/O for l logical qubits
+- **Union-Find decoder**: 100× higher accuracy than greedy algorithms
+- **Prototype**: 100 logical qubits on 5 VMK-180 FPGAs
+
+**Relevance to ruQu**: Our `ParallelFabric` uses flat parallelism. Consider hierarchical tree-grid topology for 1000+ tile scaling.
+
+### Google Below-Threshold (2025)
+**Source**: [Nature 2024](https://www.nature.com/articles/s41586-024-08449-y)
+
+Google achieved Λ = 2.14 ± 0.02 error suppression when increasing code distance by 2, with a 101-qubit distance-7 code achieving **0.143% error per cycle**.
+
+**Relevance to ruQu**: Our three-filter decision pipeline should target similar sub-0.2% false positive rates.
+
+---
+
+## 2. Hardware-Accelerated Decoding
+
+### Riverlane Collision Clustering Decoder
+**Source**: [Riverlane Blog](https://www.riverlane.com/news/introducing-the-world-s-first-low-latency-qec-experiment)
+
+| Platform | Qubits | Latency | Power |
+|----------|--------|---------|-------|
+| FPGA | 881 | 810 ns | - |
+| ASIC | 1,057 | **240 ns** | 8 mW |
+
+The ASIC fits in 0.06 mm² - suitable for cryogenic deployment.
+
+**Relevance to ruQu**: Our coherence simulation achieves 468ns P99. ASIC compilation of the hot path could reach 240ns.
+
+### QASBA: Sparse Blossom on FPGA
+**Source**: [ACM TRETS](https://dl.acm.org/doi/10.1145/3723168)
+
+- **25× performance** vs software baseline
+- **304× energy efficiency** improvement
+
+**Relevance to ruQu**: Our min-cut computation is the hot path. FPGA synthesis of `SubpolynomialMinCut` could yield similar gains.
+
+---
+
+## 3. Adaptive Syndrome Extraction
+
+### PRX Quantum (July 2025)
+**Source**: [PRX Quantum](https://doi.org/10.1103/ps3r-wf84)
+
+Adaptive syndrome extraction measures **only stabilizers likely to provide useful information**:
+
+- **10× lower logical error rates** vs non-adaptive
+- Fewer CNOT gates and physical qubits
+- Uses [[4,2,2]] concatenated with hypergraph product code
+
+**Relevance to ruQu**: This validates our coherence gate philosophy - don't process everything, focus on what matters. Consider:
+- Tracking which detectors fire frequently (already in `stim.rs`)
+- Skip syndrome processing for "quiet" regions
+- Adaptive measurement scheduling
+
+### Multi-Agent RL for QEC
+**Source**: [arXiv:2509.03974](https://arxiv.org/pdf/2509.03974)
+
+Uses **reinforcement learning bandits** to:
+- Evaluate fidelity after recovery
+- Determine when retraining is necessary
+- Optimize encoder, syndrome measurement, and recovery jointly
+
+**Relevance to ruQu**: Our `AdaptiveThresholds` uses EMA-based learning. Consider upgrading to bandit-based exploration for threshold optimization.
+
+### Window-Based Drift Estimation (Nov 2025)
+**Source**: [arXiv:2511.09491](https://arxiv.org/html/2511.09491)
+
+Estimates noise drift profiles **from syndrome data alone**, then adapts decoder parameters.
+
+**Relevance to ruQu**: Integrate drift detection into `adaptive.rs`:
+```rust
+pub fn detect_drift(&mut self, window: &[SyndromeStats]) -> Option<DriftProfile> {
+    // Detect if noise characteristics are shifting
+    // Adjust thresholds proactively
+}
+```
+
+---
+
+## 4. Mixture-of-Depths for Efficiency
+
+### MoD (DeepMind, 2024)
+**Source**: [arXiv:2404.02258](https://arxiv.org/html/2404.02258v1)
+
+- **50% FLOPs reduction** while matching dense transformer performance
+- Per-token dynamic routing (skip middle layers for "resolved" tokens)
+- Different from early-exit: tokens can skip middle layers then attend
+
+**Status**: Already implemented in `attention.rs` via `MincutDepthRouter` integration.
+
+### Mixture-of-Recursions (NeurIPS 2025)
+**Source**: [arXiv:2507.10524](https://arxiv.org/html/2507.10524v1)
+
+Combines parameter sharing + adaptive computation:
+- Reuses shared layer stack across recursion steps
+- Lightweight routers assign recursion depth per-token
+- Token-level early exiting for simple predictions
+
+**Relevance to ruQu**: Consider recursive tile processing:
+```rust
+pub fn process_recursive(&mut self, syndrome: &SyndromeDelta, max_depth: usize) -> GateDecision {
+    for depth in 0..max_depth {
+        let decision = self.process_at_depth(syndrome, depth);
+        if decision.confidence > EARLY_EXIT_THRESHOLD {
+            return decision;  // Exit early for clear cases
+        }
+    }
+    decision
+}
+```
+
+---
+
+## 5. Fusion Blossom Performance
+
+### Fusion Blossom Decoder
+**Source**: [arXiv:2305.08307](https://arxiv.org/abs/2305.08307), [GitHub](https://github.com/yuewuo/fusion-blossom)
+
+- **1 million measurement rounds/second** at d=33
+- **0.7 ms latency** in stream mode at d=21
+- **58 ns per non-trivial measurement** on 64-core machine
+- O(N) complexity for defect vertices N
+
+**Status**: Already integrated via `decoder.rs` feature. Consider:
+- Enabling parallel fusion mode in production
+- Streaming mode for real-time applications
+
+### PyMatching V2 Comparison
+PyMatching V2 achieves 5-20× single-thread speedup over Fusion Blossom. The algorithms are compatible - combining them could yield another 5-20× improvement.
+
+---
+
+## 6. Graph Neural Networks for QEC
+
+### QSeer (May 2025)
+**Source**: [arXiv:2505.06810](https://arxiv.org/abs/2505.06810)
+
+GNN for QAOA parameter prediction:
+- 6-68% improvement in approximation ratio
+- 5-10× convergence speedup
+- Supports variable-depth circuits and weighted Max-Cut
+
+**Relevance to ruQu**: Train a small GNN to predict optimal thresholds from syndrome graph structure:
+```rust
+pub struct ThresholdPredictor {
+    model: OnnxModel,  // Export trained model
+}
+
+impl ThresholdPredictor {
+    pub fn predict(&self, graph_embedding: &[f32]) -> GateThresholds {
+        // Use learned model for threshold prediction
+    }
+}
+```
+
+---
+
+## Implementation Priority Matrix
+
+| Enhancement | Impact | Effort | Priority |
+|-------------|--------|--------|----------|
+| Hierarchical tree-grid topology | High | High | P2 |
+| Drift detection in adaptive.rs | High | Medium | P1 |
+| Recursive early-exit processing | Medium | Low | P1 |
+| Bandit-based threshold exploration | Medium | Medium | P2 |
+| FPGA synthesis of min-cut | Very High | Very High | P3 |
+| GNN threshold predictor | Medium | High | P3 |
+| Streaming Fusion mode | High | Low | P1 |
+
+---
+
+## Immediate Next Steps
+
+1. **Drift Detection**: Add window-based drift estimation to `adaptive.rs`
+2. **Early-Exit Depth**: Implement confidence-based early exit in tile processing
+3. **Streaming Decoder**: Enable Fusion Blossom streaming mode for <1ms latency
+4. **Parallel Fusion**: Configure parallel fusion on 64+ core systems
+
+---
+
+## References
+
+1. DECONET: [arxiv.org/abs/2504.11805](https://arxiv.org/abs/2504.11805)
+2. Google Below-Threshold: [nature.com/articles/s41586-024-08449-y](https://www.nature.com/articles/s41586-024-08449-y)
+3. Riverlane CC Decoder: [riverlane.com](https://www.riverlane.com/news/introducing-the-world-s-first-low-latency-qec-experiment)
+4. Adaptive Syndrome Extraction: [doi.org/10.1103/ps3r-wf84](https://doi.org/10.1103/ps3r-wf84)
+5. Multi-Agent RL QEC: [arxiv.org/pdf/2509.03974](https://arxiv.org/pdf/2509.03974)
+6. Drift Estimation: [arxiv.org/html/2511.09491](https://arxiv.org/html/2511.09491)
+7. Mixture-of-Depths: [arxiv.org/html/2404.02258v1](https://arxiv.org/html/2404.02258v1)
+8. Mixture-of-Recursions: [arxiv.org/html/2507.10524v1](https://arxiv.org/html/2507.10524v1)
+9. Fusion Blossom: [arxiv.org/abs/2305.08307](https://arxiv.org/abs/2305.08307)
+10. QSeer GNN: [arxiv.org/abs/2505.06810](https://arxiv.org/abs/2505.06810)
+11. QASBA FPGA: [dl.acm.org/doi/10.1145/3723168](https://dl.acm.org/doi/10.1145/3723168)