Fix code review issues from gist analysis

[ruvnet]

Summary

This issue tracks the fixes for critical code issues identified in the external code review:
https://gist.github.com/couzic/93126a1c12b8d77651f93a7805b4bd60

Issues Identified and Fixed

1. ✅ Fabricated Benchmarks

• Problem: Hardcoded multipliers and simulated results in benchmark code
• Fix: Removed fabricated data, added disclaimers about missing comparative benchmarks

2. ✅ Fake Text Embeddings

• Problem: Hash-based embeddings that don't capture semantic meaning
• Fix: Implemented pluggable EmbeddingProvider trait with:
  • HashEmbedding (legacy, with warnings)
  • ApiEmbedding (OpenAI, Cohere, Voyage)
  • CandleEmbedding (local transformer models)

3. ✅ Incomplete GNN Training

• Problem: Unimplemented Loss::compute() and Loss::gradient() stubs
• Fix: Full implementation with MSE, CrossEntropy, BinaryCrossEntropy

4. ✅ Distance Function Bugs

• Problem: Overflow issues and asymmetric distance calculations
• Fix: Fixed dequantization formula, improved scale handling

5. ✅ Empty Transaction Tests

• Problem: 23 of 26 transaction tests were empty stubs
• Fix: Implemented 10+ critical tests for ACID, MVCC, isolation levels

Changes

• crates/ruvector-core/src/embeddings.rs (new)
• crates/ruvector-gnn/src/training.rs
• crates/ruvector-core/src/quantization.rs
• crates/ruvector-router-core/src/quantization.rs
• crates/ruvector-graph/tests/transaction_tests.rs
• docs/benchmarks/BENCHMARK_COMPARISON.md
• benchmarks/ comparison code

Published Packages

• 26 Rust crates to crates.io (v0.1.22)
• NPM packages updated to v0.1.25 with new native binaries