Medium priority improvements: overflow, NEON, HNSW layers, allocations

[ruvnet]

Priority: Medium

Collection of medium-priority improvements identified during code review.

1. Integer overflow risk in node ID (theoretical)

• Location: HNSW node allocation
• Risk: Node IDs could overflow on very large indexes (billions of vectors)
• Fix: Use checked arithmetic or u64

2. Missing NEON Manhattan distance

• Location: src/distance/simd.rs
• Issue: Manhattan (L1) distance not implemented for ARM NEON
• Impact: Falls back to scalar on ARM

3. Hardcoded 32-layer max in HNSW

• Location: HNSW index implementation
• Issue: Maximum 32 layers hardcoded
• Impact: May limit index depth for very large datasets
• Fix: Make configurable or use dynamic allocation

4. Memory allocation in search hot path

• Location: HNSW/IVFFlat search functions
• Issue: .clone() calls during search allocate memory
• Impact: GC pressure, cache misses
• Fix: Use arena allocator or pre-allocated buffers

5. Fragile typmod array parsing

• Location: Vector type parsing
• Issue: Manual parsing of dimension from typmod
• Impact: Could fail on edge cases
• Fix: Use proper PostgreSQL typmod utilities