Skip to content

feat(prime-radiant): Advanced Mathematical Frameworks + fix(router): VectorDb Deadlock (#133)#132

Merged
ruvnet merged 20 commits intomainfrom
feat/prime-radiant-advanced-math
Jan 24, 2026
Merged

feat(prime-radiant): Advanced Mathematical Frameworks + fix(router): VectorDb Deadlock (#133)#132
ruvnet merged 20 commits intomainfrom
feat/prime-radiant-advanced-math

Conversation

@ruvnet
Copy link
Owner

@ruvnet ruvnet commented Jan 23, 2026

Summary

This PR includes:

🐛 BUG FIX: VectorDb Deadlock (#133)

Fixed HNSW index deadlock that caused VectorDb.insert() to hang on second call.

  • Root cause: Calling search_knn_internal() while holding write locks on graph and entry_point
  • parking_lot::RwLock is NOT reentrant, causing deadlock when acquiring read locks
  • Fix: Release write locks before calling search_knn_internal(), re-acquire after
  • Added regression tests: test_hnsw_multiple_inserts_no_deadlock, test_hnsw_concurrent_inserts

Version Bump:

  • Rust workspace: 2.0.02.0.1
  • npm @ruvector/router: 0.1.250.1.26

🧮 Prime-Radiant Mathematical Frameworks

Implements 6 cutting-edge mathematical frameworks for the Prime-Radiant coherence engine, providing rigorous foundations for AI interpretability, memory coherence, and neural network analysis.

Mathematical Directions Implemented

  1. Sheaf Cohomology (src/cohomology/)

    • Coboundary operators: δ: C^n → C^{n+1}
    • Sheaf Laplacian: E(S) = Σ w_e ||ρ_u(x_u) - ρ_v(x_v)||²
    • Sheaf Neural Networks for coherent attention
  2. Category Theory & Topos (src/category/)

    • Core Category trait with identity/composition
    • Functors, natural transformations
    • Topos with subobject classifiers
  3. Homotopy Type Theory (src/hott/)

    • Dependent types (Pi, Sigma, Id)
    • Path types and J-eliminator
    • Univalence axiom: (A ≃ B) ≃ (A = B)
  4. Spectral Invariants (src/spectral/)

    • Lanczos eigensolver for sparse matrices
    • Cheeger bounds: λ₂/2 ≤ h(G) ≤ √(2λ₂)
  5. Causal Abstraction Networks (src/causal/)

    • Structural Causal Models (SCMs)
    • Do-calculus engine with 3 rules
  6. Quantum/Algebraic Topology (src/quantum/)

    • Quantum states (GHZ, W states)
    • Persistent homology & Betti numbers

Test plan

  • cargo test -p ruvector-router-core - Index tests pass
  • cargo test - All 191+ tests passing
  • WASM build succeeds

Fixes #133

🤖 Generated with Claude Code

Reuven and others added 19 commits January 22, 2026 11:24
…ence engine

Add comprehensive architecture documentation for ruvector-coherence crate:

- ADR-014: Sheaf Laplacian-based coherence witnessing architecture
  - Universal coherence object with domain-agnostic interpretation
  - 5-layer architecture (Application → Gate → Computation → Governance → Storage)
  - 4-tier compute ladder (Reflex → Retrieval → Heavy → Human)
  - Full ruvector ecosystem integration (10+ crates)
  - 15 internal architectural decisions

- DDD: Domain-Driven Design with 10 bounded contexts
  - Tile Fabric (cognitum-gate-kernel)
  - Adaptive Learning (sona)
  - Neural Gating (ruvector-nervous-system)
  - Learned Restriction Maps (ruvector-gnn)
  - Hyperbolic Coherence (ruvector-hyperbolic-hnsw)
  - Incoherence Isolation (ruvector-mincut)
  - Attention-Weighted Coherence (ruvector-attention)
  - Distributed Consensus (ruvector-raft)

Key concept: "This is not prediction. It is a continuously updated field
of coherence that shows where action is safe and where action must stop."

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implement the complete Prime-Radiant crate based on ADR-014:

Core Modules:
- substrate/: SheafGraph, SheafNode, SheafEdge, RestrictionMap (SIMD-optimized)
- coherence/: CoherenceEngine, energy computation, spectral drift detection
- governance/: PolicyBundle, WitnessRecord, LineageRecord (Blake3 hashing)
- execution/: CoherenceGate, ComputeLane, ActionExecutor

Ecosystem Integrations (feature-gated):
- tiles/: cognitum-gate-kernel 256-tile WASM fabric adapter
- sona_tuning/: Adaptive threshold learning with EWC++
- neural_gate/: Biologically-inspired gating with HDC encoding
- learned_rho/: GNN-based learned restriction maps
- attention/: Topology-gated attention, MoE routing, PDE diffusion
- distributed/: Raft-based multi-node coherence

Testing:
- 138 tests (integration, property-based, chaos)
- 8 benchmarks covering ADR-014 performance targets

Stats: 91 files, ~30K lines of Rust code

"This is not prediction. It is a continuously updated field of coherence
that shows where action is safe and where action must stop."

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add coherence-gated LLM inference architecture diagram
- Add 5 integration modules with code examples:
  - SheafCoherenceValidator (replaces heuristic scoring)
  - UnifiedWitnessLog (merged audit trail)
  - PatternToRestrictionBridge (ReasoningBank → learned ρ)
  - MemoryCoherenceLayer (context as sheaf nodes)
  - CoherenceConfidence (energy → confidence mapping)
- Add 7 integration ADRs (ADR-CE-016 through ADR-CE-022)
- Add ruvllm to crate integration matrix and dependencies
- Add 4 LLM-specific benefits to consequences
- Add ruvllm feature flag

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Create detailed ADR files for all internal coherence engine decisions:

Core Architecture (ADR-CE-001 to ADR-CE-008):
- 001: Sheaf Laplacian defines coherence witness
- 002: Incremental computation with stored residuals
- 003: PostgreSQL + ruvector hybrid storage
- 004: Signed event log with deterministic replay
- 005: First-class governance objects
- 006: Coherence gate controls compute ladder
- 007: Thresholds auto-tuned from traces
- 008: Multi-tenant isolation boundaries

Universal Coherence (ADR-CE-009 to ADR-CE-015):
- 009: Single coherence object (one math, many interpretations)
- 010: Domain-agnostic nodes and edges
- 011: Residual = contradiction energy
- 012: Gate = refusal mechanism with witness
- 013: Not prediction (coherence field, not forecasting)
- 014: Reflex lane default (most ops stay fast)
- 015: Adapt without losing control

RuvLLM Integration (ADR-CE-016 to ADR-CE-022):
- 016: CoherenceValidator uses sheaf energy
- 017: Unified audit trail (WitnessLog + governance)
- 018: Pattern-to-restriction bridge (ReasoningBank)
- 019: Memory as nodes (agentic, working, episodic)
- 020: Confidence from energy (sigmoid mapping)
- 021: Shared SONA between ruvllm and prime-radiant
- 022: Failure learning (ErrorPatternLearner → ρ maps)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implement complete Prime-Radiant + RuvLLM integration per ADR-CE-016 through ADR-CE-022:

Core Integration Modules:
- coherence_validator.rs: SheafCoherenceValidator using sheaf energy
- witness_log.rs: UnifiedWitnessLog with hash chain for tamper evidence
- pattern_bridge.rs: PatternToRestrictionBridge learning from verdicts
- memory_layer.rs: MemoryCoherenceLayer tracking context as sheaf nodes
- confidence.rs: CoherenceConfidence with sigmoid energy→confidence mapping

Supporting Infrastructure:
- mod.rs: Public API, re-exports, convenience constructors
- error.rs: Comprehensive error types for each ADR
- config.rs: LlmCoherenceConfig, thresholds, policies
- gate.rs: LlmCoherenceGate high-level interface
- adapter.rs: RuvLlmAdapter bridging type systems
- bridge.rs: PolicyBridge, SonaBridge for synchronization
- witness.rs: WitnessAdapter for correlation
- traits.rs: Trait definitions for loose coupling

Testing:
- 22 integration tests covering all modules
- Self-contained mock implementations
- Feature-gated with #[cfg(feature = "ruvllm")]

Feature Flags:
- ruvllm feature in Cargo.toml
- Optional dependency on ruvllm crate
- Added to "full" feature set

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add user-friendly documentation covering:
- Introduction explaining coherence vs confidence
- Core concepts (coherence field, compute ladder)
- Features overview (engine, governance, RuvLLM integration)
- Quick start code examples:
  - Basic coherence check
  - LLM response validation
  - Memory consistency tracking
  - Confidence from energy
- Application tiers (today, near-term, future)
- Domain examples (AI, finance, medical, robotics, security)
- Feature flags reference
- Performance targets
- Architecture diagram

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Propose novel low-latency transformer architecture using coherence energy:

Core Innovation:
- Route tokens to compute lanes based on coherence energy, not confidence
- Sparse attention using residual energy (skip coherent pairs)
- Early exit when energy converges (not confidence threshold)
- Restriction maps replace QKV projections

Architecture:
- Lane 0 (Reflex): 1-2 layers, local attention, <0.1ms
- Lane 1 (Standard): 6 layers, sparse sheaf attention, ~1ms
- Lane 2 (Deep): 12+ layers, full + MoE, ~5ms
- Lane 3 (Escalate): Return uncertainty

Performance Targets:
- 5-10x latency reduction (10ms → 1-2ms for 128 tokens)
- 2.5x memory reduction
- <5% quality degradation
- Provable coherence bound on output

Mathematical Foundation:
- Attention weight ∝ exp(-β × residual_energy)
- Token routing via E(t) = Σ w_e ||ρ_t(x) - ρ_ctx(x)||²
- Early exit when ΔE < ε (energy converged)

Target: ruvector-attention crate with sheaf/ and coherence_gated/ modules

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Complete implementation of Prime-Radiant coherence engine and
Coherence-Gated Transformer (CGT) sheaf attention module.

Core Features:
- Sheaf Laplacian energy computation with restriction maps
- 4-lane compute ladder (Reflex/Retrieval/Heavy/Human)
- Cryptographic witness chains for audit trails
- Policy bundles with multi-party approval

Storage Backends:
- InMemoryStorage with KNN search
- FileStorage with Write-Ahead Logging (WAL)
- PostgresStorage with full schema (feature-gated)
- HybridStorage combining file + optional PostgreSQL

CGT Sheaf Attention (ruvector-attention):
- RestrictionMap with residual/energy computation
- SheafAttention layer: A_ij = exp(-β×E_ij)/Z
- TokenRouter with compute lane routing
- SparseResidualAttention with energy-based masking
- EarlyExit with energy convergence detection

Performance Optimizations:
- Zero-allocation hot paths (apply_into, compute_residual_norm_sq)
- SIMD-friendly 4-way unrolled loops
- Branchless lane routing
- Pre-allocated buffers for batch operations

RuvLLM Integration:
- SheafCoherenceValidator for LLM response validation
- UnifiedWitnessLog linking inference + coherence
- MemoryCoherenceLayer for contradiction detection
- CoherenceConfidence for interpretable uncertainty

Tests: 202 passing in ruvector-attention, 180+ in prime-radiant

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
…nchmarks

GPU Acceleration (wgpu-rs):
- GpuCoherenceEngine with automatic CPU fallback
- GpuDevice: adapter/device management with high-perf selection
- GpuDispatcher: kernel execution with pipeline caching and buffer pooling
- GpuBufferManager: typed buffer management with pooling
- Compute kernels: residuals, energy reduction, sheaf attention, token routing

WGSL Compute Shaders (6 files, 1,412 lines):
- compute_residuals.wgsl: parallel edge residual computation
- compute_energy.wgsl: two-phase parallel reduction
- sheaf_attention.wgsl: energy-based attention weights A_ij = exp(-beta * E_ij)
- token_routing.wgsl: branchless lane assignment
- sparse_mask.wgsl: sparse attention mask generation
- types.wgsl: shared GPU struct definitions

SIMD Optimizations (wide crate):
- Runtime CPU feature detection (AVX2, AVX-512, SSE4.2, NEON)
- f32x8 vectorized operations
- simd/vectors.rs: dot_product_simd, norm_squared_simd, subtract_simd
- simd/matrix.rs: matmul_simd, matvec_simd, transpose_simd
- simd/energy.rs: batch_residuals_simd, weighted_energy_sum_simd
- 38 unit tests verifying SIMD correctness

Benchmarks (criterion):
- coherence_benchmarks.rs: core operations, graph scaling
- simd_benchmarks.rs: SIMD vs naive comparisons
- gpu_benchmarks.rs: CPU vs GPU performance

Tests:
- 18 GPU coherence tests (16 active, 2 perf ignored)
- GPU-CPU consistency within 1% relative error
- Error handling and fallback verification

README improvements:
- "What Prime-Radiant is NOT" section
- Concrete numeric example with arithmetic
- Flagship LLM hallucination refusal walkthrough
- Infrastructure positioning

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
SIMD Optimizations:
- Replace element-by-element load_f32x8 with try_into for direct memory copy
- Fix redundant SIMD comparisons in lane assignment (compute masks once, use blend)
- Apply across vectors.rs, matrix.rs, and energy.rs

Core Computation Patterns:
- Replace i % 4 modulo with chunks_exact() for proper auto-vectorization
- Fix edge.rs: residual_norm_squared, residual_with_energy
- Fix node.rs: norm_squared, dot product

Graph API:
- Add get_node_ref() for zero-copy node access via DashMap reference
- Add with_node() closure API for efficient read-only operations

Benchmark findings:
- Incremental updates meet target (<100us): 59us actual
- Linear O(n) scaling confirmed
- Further SIMD/parallelization needed for <1us/edge target

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
…ad-local scratch

Performance optimizations for Prime-Radiant coherence engine:

CSR Sparse Matrix (restriction.rs):
- Full CsrMatrix struct with row_ptr, col_indices, values
- COO to CSR conversion with from_coo() and from_coo_arrays()
- Zero-allocation matvec_into() and matvec_add_into()
- SIMD-friendly 4-element loop unrolling
- 13 new tests covering all CSR operations

GPU Buffer Pre-allocation (engine.rs, kernels.rs):
- Pre-allocated params, energy_params, partial_sums, staging buffers
- Zero per-frame allocations in compute_energy()
- New create_bind_group_raw() methods for raw buffer references
- CSR matrix support in convert_restriction_map()

Thread-Local Scratch Buffers (edge.rs):
- EdgeScratch struct with 3 reusable Vec<f32> buffers
- thread_local! SCRATCH for zero-allocation hot paths
- residual_norm_squared_no_alloc() and weighted_residual_energy_no_alloc()
- 7 new tests for allocation-free energy computation

WGSL Vec4 Optimization (compute_residuals.wgsl):
- vec4-based processing loop with dot(r_vec, r_vec)
- store_residuals flag in GpuParams struct
- ~4x GPU throughput improvement

README Updates:
- Root README: 40 attention mechanisms, Prime-Radiant section, CGT Sheaf Attention
- WASM README: CGT Sheaf Attention API documentation

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- prime-radiant: Enhanced description, keywords, categories
- ruvector-attention-wasm: Add version to path dep, SEO keywords
- package.json: 23 keywords, better description, engines config

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
…eferences

- Add prime-radiant to Quantum Coherence section (sheaf Laplacian AI safety)
- Add ruvector-attention-wasm to npm WASM packages (Flash, MoE, Hyperbolic, CGT)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Comprehensive implementation of cutting-edge mathematical foundations:

## Modules Implemented

1. **Sheaf Cohomology** (10 files)
   - Coboundary operator, Cohomology groups, Betti numbers
   - Sheaf Laplacian, Obstruction detection, Diffusion
   - Sheaf Neural Networks with CohomologyPooling

2. **Category Theory/Topos** (12 files)
   - Category trait, Functors, Natural transformations
   - Topos with SubobjectClassifier, InternalLogic
   - 2-Category with Mac Lane coherence (pentagon/triangle)
   - BeliefTopos for probabilistic reasoning

3. **Homotopy Type Theory** (8 files)
   - Type/Term AST with Pi, Sigma, Identity types
   - Path operations, J-eliminator, Transport
   - Univalence axiom, Bidirectional type checker
   - Coherence as paths between belief states

4. **Spectral Invariants** (8 files)
   - Lanczos eigensolver for sparse matrices
   - Cheeger inequality bounds and sweep algorithm
   - Spectral clustering with k-means++
   - Collapse prediction and early warning system

5. **Causal Abstraction** (7 files)
   - Structural Causal Models with do-calculus
   - D-separation (Bayes Ball), Topological ordering
   - Counterfactuals: ATE, ITE, NDE, NIE
   - Causal abstraction verification

6. **Quantum/Algebraic Topology** (10 files)
   - Quantum states, Density matrices, Channels
   - Simplicial complexes, Persistent homology
   - Topological codes (surface, toric, stabilizer)
   - Structure-preserving quantum encodings

## Supporting Infrastructure

- **Security Module**: 17 issues fixed, path traversal prevention
- **WASM Bindings**: 6 engines with TypeScript definitions
- **Benchmarks**: 4,762 lines of criterion benchmarks
- **Documentation**: 6 ADRs + DDD domain model (3,141 lines)
- **Tests**: 191+ tests passing

## Mathematical Foundations

- Sheaf Laplacian: E(S) = Σ w_e ||ρ_u(x_u) - ρ_v(x_v)||²
- Cheeger inequality: λ₂/2 ≤ h(G) ≤ √(2λ₂)
- Univalence: (A ≃ B) ≃ (A = B)
- Do-calculus: P(Y|do(X)) identification

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The insert() method was holding write locks on graph and entry_point
while calling search_knn_internal(), which tries to acquire read locks
on the same RwLocks. Since parking_lot::RwLock is NOT reentrant, this
caused a deadlock on the second insert.

Fix: Release all locks before calling search_knn_internal(), then
re-acquire for modifications.

Added regression tests:
- test_hnsw_multiple_inserts_no_deadlock
- test_hnsw_concurrent_inserts

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Rust workspace: 2.0.0 -> 2.0.1
- npm @ruvector/router: 0.1.25 -> 0.1.26
- npm platform packages: -> 0.1.26
- Added darwin-x64 to optional dependencies

Contains fix for HNSW deadlock issue #133

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@ruvnet ruvnet changed the title feat(prime-radiant): Advanced Mathematical Frameworks for AI Interpretability feat(prime-radiant): Advanced Mathematical Frameworks + fix(router): VectorDb Deadlock (#133) Jan 24, 2026
@ruvnet ruvnet merged commit 6c50685 into main Jan 24, 2026
15 checks passed
@ruvnet ruvnet deleted the feat/prime-radiant-advanced-math branch January 24, 2026 17:31
ruvnet added a commit that referenced this pull request Feb 20, 2026
…VectorDb Deadlock (#133) (#132)

* docs(coherence-engine): add ADR-014 and DDD for sheaf Laplacian coherence engine

Add comprehensive architecture documentation for ruvector-coherence crate:

- ADR-014: Sheaf Laplacian-based coherence witnessing architecture
  - Universal coherence object with domain-agnostic interpretation
  - 5-layer architecture (Application → Gate → Computation → Governance → Storage)
  - 4-tier compute ladder (Reflex → Retrieval → Heavy → Human)
  - Full ruvector ecosystem integration (10+ crates)
  - 15 internal architectural decisions

- DDD: Domain-Driven Design with 10 bounded contexts
  - Tile Fabric (cognitum-gate-kernel)
  - Adaptive Learning (sona)
  - Neural Gating (ruvector-nervous-system)
  - Learned Restriction Maps (ruvector-gnn)
  - Hyperbolic Coherence (ruvector-hyperbolic-hnsw)
  - Incoherence Isolation (ruvector-mincut)
  - Attention-Weighted Coherence (ruvector-attention)
  - Distributed Consensus (ruvector-raft)

Key concept: "This is not prediction. It is a continuously updated field
of coherence that shows where action is safe and where action must stop."

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* feat(prime-radiant): implement sheaf Laplacian coherence engine

Implement the complete Prime-Radiant crate based on ADR-014:

Core Modules:
- substrate/: SheafGraph, SheafNode, SheafEdge, RestrictionMap (SIMD-optimized)
- coherence/: CoherenceEngine, energy computation, spectral drift detection
- governance/: PolicyBundle, WitnessRecord, LineageRecord (Blake3 hashing)
- execution/: CoherenceGate, ComputeLane, ActionExecutor

Ecosystem Integrations (feature-gated):
- tiles/: cognitum-gate-kernel 256-tile WASM fabric adapter
- sona_tuning/: Adaptive threshold learning with EWC++
- neural_gate/: Biologically-inspired gating with HDC encoding
- learned_rho/: GNN-based learned restriction maps
- attention/: Topology-gated attention, MoE routing, PDE diffusion
- distributed/: Raft-based multi-node coherence

Testing:
- 138 tests (integration, property-based, chaos)
- 8 benchmarks covering ADR-014 performance targets

Stats: 91 files, ~30K lines of Rust code

"This is not prediction. It is a continuously updated field of coherence
that shows where action is safe and where action must stop."

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* docs(adr): add RuvLLM integration to ADR-014 v0.4

- Add coherence-gated LLM inference architecture diagram
- Add 5 integration modules with code examples:
  - SheafCoherenceValidator (replaces heuristic scoring)
  - UnifiedWitnessLog (merged audit trail)
  - PatternToRestrictionBridge (ReasoningBank → learned ρ)
  - MemoryCoherenceLayer (context as sheaf nodes)
  - CoherenceConfidence (energy → confidence mapping)
- Add 7 integration ADRs (ADR-CE-016 through ADR-CE-022)
- Add ruvllm to crate integration matrix and dependencies
- Add 4 LLM-specific benefits to consequences
- Add ruvllm feature flag

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* docs(adr): add 22 coherence engine internal ADRs

Create detailed ADR files for all internal coherence engine decisions:

Core Architecture (ADR-CE-001 to ADR-CE-008):
- 001: Sheaf Laplacian defines coherence witness
- 002: Incremental computation with stored residuals
- 003: PostgreSQL + ruvector hybrid storage
- 004: Signed event log with deterministic replay
- 005: First-class governance objects
- 006: Coherence gate controls compute ladder
- 007: Thresholds auto-tuned from traces
- 008: Multi-tenant isolation boundaries

Universal Coherence (ADR-CE-009 to ADR-CE-015):
- 009: Single coherence object (one math, many interpretations)
- 010: Domain-agnostic nodes and edges
- 011: Residual = contradiction energy
- 012: Gate = refusal mechanism with witness
- 013: Not prediction (coherence field, not forecasting)
- 014: Reflex lane default (most ops stay fast)
- 015: Adapt without losing control

RuvLLM Integration (ADR-CE-016 to ADR-CE-022):
- 016: CoherenceValidator uses sheaf energy
- 017: Unified audit trail (WitnessLog + governance)
- 018: Pattern-to-restriction bridge (ReasoningBank)
- 019: Memory as nodes (agentic, working, episodic)
- 020: Confidence from energy (sigmoid mapping)
- 021: Shared SONA between ruvllm and prime-radiant
- 022: Failure learning (ErrorPatternLearner → ρ maps)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* feat(prime-radiant): implement RuvLLM integration layer (ADR-014 v0.4)

Implement complete Prime-Radiant + RuvLLM integration per ADR-CE-016 through ADR-CE-022:

Core Integration Modules:
- coherence_validator.rs: SheafCoherenceValidator using sheaf energy
- witness_log.rs: UnifiedWitnessLog with hash chain for tamper evidence
- pattern_bridge.rs: PatternToRestrictionBridge learning from verdicts
- memory_layer.rs: MemoryCoherenceLayer tracking context as sheaf nodes
- confidence.rs: CoherenceConfidence with sigmoid energy→confidence mapping

Supporting Infrastructure:
- mod.rs: Public API, re-exports, convenience constructors
- error.rs: Comprehensive error types for each ADR
- config.rs: LlmCoherenceConfig, thresholds, policies
- gate.rs: LlmCoherenceGate high-level interface
- adapter.rs: RuvLlmAdapter bridging type systems
- bridge.rs: PolicyBridge, SonaBridge for synchronization
- witness.rs: WitnessAdapter for correlation
- traits.rs: Trait definitions for loose coupling

Testing:
- 22 integration tests covering all modules
- Self-contained mock implementations
- Feature-gated with #[cfg(feature = "ruvllm")]

Feature Flags:
- ruvllm feature in Cargo.toml
- Optional dependency on ruvllm crate
- Added to "full" feature set

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* docs(prime-radiant): add comprehensive README with examples

Add user-friendly documentation covering:
- Introduction explaining coherence vs confidence
- Core concepts (coherence field, compute ladder)
- Features overview (engine, governance, RuvLLM integration)
- Quick start code examples:
  - Basic coherence check
  - LLM response validation
  - Memory consistency tracking
  - Confidence from energy
- Application tiers (today, near-term, future)
- Domain examples (AI, finance, medical, robotics, security)
- Feature flags reference
- Performance targets
- Architecture diagram

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* docs(adr): add ADR-015 Coherence-Gated Transformer (Sheaf Attention)

Propose novel low-latency transformer architecture using coherence energy:

Core Innovation:
- Route tokens to compute lanes based on coherence energy, not confidence
- Sparse attention using residual energy (skip coherent pairs)
- Early exit when energy converges (not confidence threshold)
- Restriction maps replace QKV projections

Architecture:
- Lane 0 (Reflex): 1-2 layers, local attention, <0.1ms
- Lane 1 (Standard): 6 layers, sparse sheaf attention, ~1ms
- Lane 2 (Deep): 12+ layers, full + MoE, ~5ms
- Lane 3 (Escalate): Return uncertainty

Performance Targets:
- 5-10x latency reduction (10ms → 1-2ms for 128 tokens)
- 2.5x memory reduction
- <5% quality degradation
- Provable coherence bound on output

Mathematical Foundation:
- Attention weight ∝ exp(-β × residual_energy)
- Token routing via E(t) = Σ w_e ||ρ_t(x) - ρ_ctx(x)||²
- Early exit when ΔE < ε (energy converged)

Target: ruvector-attention crate with sheaf/ and coherence_gated/ modules

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* feat(prime-radiant): implement coherence engine with CGT attention

Complete implementation of Prime-Radiant coherence engine and
Coherence-Gated Transformer (CGT) sheaf attention module.

Core Features:
- Sheaf Laplacian energy computation with restriction maps
- 4-lane compute ladder (Reflex/Retrieval/Heavy/Human)
- Cryptographic witness chains for audit trails
- Policy bundles with multi-party approval

Storage Backends:
- InMemoryStorage with KNN search
- FileStorage with Write-Ahead Logging (WAL)
- PostgresStorage with full schema (feature-gated)
- HybridStorage combining file + optional PostgreSQL

CGT Sheaf Attention (ruvector-attention):
- RestrictionMap with residual/energy computation
- SheafAttention layer: A_ij = exp(-β×E_ij)/Z
- TokenRouter with compute lane routing
- SparseResidualAttention with energy-based masking
- EarlyExit with energy convergence detection

Performance Optimizations:
- Zero-allocation hot paths (apply_into, compute_residual_norm_sq)
- SIMD-friendly 4-way unrolled loops
- Branchless lane routing
- Pre-allocated buffers for batch operations

RuvLLM Integration:
- SheafCoherenceValidator for LLM response validation
- UnifiedWitnessLog linking inference + coherence
- MemoryCoherenceLayer for contradiction detection
- CoherenceConfidence for interpretable uncertainty

Tests: 202 passing in ruvector-attention, 180+ in prime-radiant

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* feat(prime-radiant): add GPU acceleration, SIMD optimizations, and benchmarks

GPU Acceleration (wgpu-rs):
- GpuCoherenceEngine with automatic CPU fallback
- GpuDevice: adapter/device management with high-perf selection
- GpuDispatcher: kernel execution with pipeline caching and buffer pooling
- GpuBufferManager: typed buffer management with pooling
- Compute kernels: residuals, energy reduction, sheaf attention, token routing

WGSL Compute Shaders (6 files, 1,412 lines):
- compute_residuals.wgsl: parallel edge residual computation
- compute_energy.wgsl: two-phase parallel reduction
- sheaf_attention.wgsl: energy-based attention weights A_ij = exp(-beta * E_ij)
- token_routing.wgsl: branchless lane assignment
- sparse_mask.wgsl: sparse attention mask generation
- types.wgsl: shared GPU struct definitions

SIMD Optimizations (wide crate):
- Runtime CPU feature detection (AVX2, AVX-512, SSE4.2, NEON)
- f32x8 vectorized operations
- simd/vectors.rs: dot_product_simd, norm_squared_simd, subtract_simd
- simd/matrix.rs: matmul_simd, matvec_simd, transpose_simd
- simd/energy.rs: batch_residuals_simd, weighted_energy_sum_simd
- 38 unit tests verifying SIMD correctness

Benchmarks (criterion):
- coherence_benchmarks.rs: core operations, graph scaling
- simd_benchmarks.rs: SIMD vs naive comparisons
- gpu_benchmarks.rs: CPU vs GPU performance

Tests:
- 18 GPU coherence tests (16 active, 2 perf ignored)
- GPU-CPU consistency within 1% relative error
- Error handling and fallback verification

README improvements:
- "What Prime-Radiant is NOT" section
- Concrete numeric example with arithmetic
- Flagship LLM hallucination refusal walkthrough
- Infrastructure positioning

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* perf(prime-radiant): optimize SIMD and core computation patterns

SIMD Optimizations:
- Replace element-by-element load_f32x8 with try_into for direct memory copy
- Fix redundant SIMD comparisons in lane assignment (compute masks once, use blend)
- Apply across vectors.rs, matrix.rs, and energy.rs

Core Computation Patterns:
- Replace i % 4 modulo with chunks_exact() for proper auto-vectorization
- Fix edge.rs: residual_norm_squared, residual_with_energy
- Fix node.rs: norm_squared, dot product

Graph API:
- Add get_node_ref() for zero-copy node access via DashMap reference
- Add with_node() closure API for efficient read-only operations

Benchmark findings:
- Incremental updates meet target (<100us): 59us actual
- Linear O(n) scaling confirmed
- Further SIMD/parallelization needed for <1us/edge target

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* perf(prime-radiant): add CSR sparse matrix, GPU buffer prealloc, thread-local scratch

Performance optimizations for Prime-Radiant coherence engine:

CSR Sparse Matrix (restriction.rs):
- Full CsrMatrix struct with row_ptr, col_indices, values
- COO to CSR conversion with from_coo() and from_coo_arrays()
- Zero-allocation matvec_into() and matvec_add_into()
- SIMD-friendly 4-element loop unrolling
- 13 new tests covering all CSR operations

GPU Buffer Pre-allocation (engine.rs, kernels.rs):
- Pre-allocated params, energy_params, partial_sums, staging buffers
- Zero per-frame allocations in compute_energy()
- New create_bind_group_raw() methods for raw buffer references
- CSR matrix support in convert_restriction_map()

Thread-Local Scratch Buffers (edge.rs):
- EdgeScratch struct with 3 reusable Vec<f32> buffers
- thread_local! SCRATCH for zero-allocation hot paths
- residual_norm_squared_no_alloc() and weighted_residual_energy_no_alloc()
- 7 new tests for allocation-free energy computation

WGSL Vec4 Optimization (compute_residuals.wgsl):
- vec4-based processing loop with dot(r_vec, r_vec)
- store_residuals flag in GpuParams struct
- ~4x GPU throughput improvement

README Updates:
- Root README: 40 attention mechanisms, Prime-Radiant section, CGT Sheaf Attention
- WASM README: CGT Sheaf Attention API documentation

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* chore: SEO optimize package metadata for crates.io and npm

- prime-radiant: Enhanced description, keywords, categories
- ruvector-attention-wasm: Add version to path dep, SEO keywords
- package.json: 23 keywords, better description, engines config

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* chore(hyperbolic-hnsw): SEO optimize for crates.io publish

* chore(prime-radiant): add version numbers to path dependencies for crates.io publish

* fix(prime-radiant): shorten keyword for crates.io compliance

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* docs(readme): add prime-radiant and ruvector-attention-wasm package references

- Add prime-radiant to Quantum Coherence section (sheaf Laplacian AI safety)
- Add ruvector-attention-wasm to npm WASM packages (Flash, MoE, Hyperbolic, CGT)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* feat(prime-radiant): implement 6 advanced mathematical frameworks

Comprehensive implementation of cutting-edge mathematical foundations:

## Modules Implemented

1. **Sheaf Cohomology** (10 files)
   - Coboundary operator, Cohomology groups, Betti numbers
   - Sheaf Laplacian, Obstruction detection, Diffusion
   - Sheaf Neural Networks with CohomologyPooling

2. **Category Theory/Topos** (12 files)
   - Category trait, Functors, Natural transformations
   - Topos with SubobjectClassifier, InternalLogic
   - 2-Category with Mac Lane coherence (pentagon/triangle)
   - BeliefTopos for probabilistic reasoning

3. **Homotopy Type Theory** (8 files)
   - Type/Term AST with Pi, Sigma, Identity types
   - Path operations, J-eliminator, Transport
   - Univalence axiom, Bidirectional type checker
   - Coherence as paths between belief states

4. **Spectral Invariants** (8 files)
   - Lanczos eigensolver for sparse matrices
   - Cheeger inequality bounds and sweep algorithm
   - Spectral clustering with k-means++
   - Collapse prediction and early warning system

5. **Causal Abstraction** (7 files)
   - Structural Causal Models with do-calculus
   - D-separation (Bayes Ball), Topological ordering
   - Counterfactuals: ATE, ITE, NDE, NIE
   - Causal abstraction verification

6. **Quantum/Algebraic Topology** (10 files)
   - Quantum states, Density matrices, Channels
   - Simplicial complexes, Persistent homology
   - Topological codes (surface, toric, stabilizer)
   - Structure-preserving quantum encodings

## Supporting Infrastructure

- **Security Module**: 17 issues fixed, path traversal prevention
- **WASM Bindings**: 6 engines with TypeScript definitions
- **Benchmarks**: 4,762 lines of criterion benchmarks
- **Documentation**: 6 ADRs + DDD domain model (3,141 lines)
- **Tests**: 191+ tests passing

## Mathematical Foundations

- Sheaf Laplacian: E(S) = Σ w_e ||ρ_u(x_u) - ρ_v(x_v)||²
- Cheeger inequality: λ₂/2 ≤ h(G) ≤ √(2λ₂)
- Univalence: (A ≃ B) ≃ (A = B)
- Do-calculus: P(Y|do(X)) identification

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* fix(router-core): resolve HNSW index deadlock on second insert (#133)

The insert() method was holding write locks on graph and entry_point
while calling search_knn_internal(), which tries to acquire read locks
on the same RwLocks. Since parking_lot::RwLock is NOT reentrant, this
caused a deadlock on the second insert.

Fix: Release all locks before calling search_knn_internal(), then
re-acquire for modifications.

Added regression tests:
- test_hnsw_multiple_inserts_no_deadlock
- test_hnsw_concurrent_inserts

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* chore: bump versions for v2.0.1 release

- Rust workspace: 2.0.0 -> 2.0.1
- npm @ruvector/router: 0.1.25 -> 0.1.26
- npm platform packages: -> 0.1.26
- Added darwin-x64 to optional dependencies

Contains fix for HNSW deadlock issue #133

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

---------

Co-authored-by: Reuven <cohen@ruv-mac-mini.local>
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

VectorDb.insert() deadlocks on second call - HNSW RwLock not reentrant

1 participant