Feat: blockwise quantization recall restore

[dev07060]

PR Description: Restore HNSW Index Recall & Implement Block-wise Quantization

This PR restores the semantic search recall on the balanced wiki evaluation dataset (recovering from a degraded state back to 1.000) and integrates a backward-compatible, block-wise scalar quantization (Q8_0 style) pipeline.

---

1. Core Problem & Resolution

A. Compound Distortion Loop (HNSW Recall Recovery)

• Issue: When the vector_quant_i8 feature was active, index rebuilding fetched the lossy i8 quantized vectors and dequantized them in memory to construct the $f32$ HNSW graph. This caused cumulative quantization distortion in HNSW's topology, damaging search paths.
• Resolution: Re-routed HNSW index rebuilding in both source_rag.rs and simple_rag.rs to load the original high-precision $f32$ embeddings (chunks.embedding) directly from the database.

B. Outlier Vulnerability (Block-wise Quantization)

• Issue: Uniform 8-bit scalar quantization calculated a single scale factor across all 768 dimensions. Global outliers compressed the dynamic range of all other dimensions, flattening meaningful vector resolutions.
• Resolution: Implemented Block-wise Quantization (GGML Q8_0 style). Vectors are partitioned into 24 blocks of 32 elements, each with an independent scale factor.

C. Zero-Schema Database Migration & Backward Compatibility

• Issue: Adding database columns for block scales would lock sqlite databases on mobile clients or break existing setups.
• Resolution: Packed scales directly into the existing chunks.embedding_i8 blob (864 bytes for a 768-dim vector). The similarity kernel checks the blob's size at runtime:
  • 768 bytes: Fallback to legacy uniform cosine math using the legacy scale.
  • 864 bytes: Parses packed scales per block for block-wise cosine dot products.

---

2. File-by-File Changes

Rust Core (rag_engine_flutter / rust_builder)

• vector_quant.rs:
  • Redefined quantize_f32_to_u8_blob under #[cfg(feature = "vector_quant_i8")] to output the 864-byte packed block-wise format.
  • Added QueryQ8 struct for precomputing block-wise quantized representations of search queries.
  • Added cosine_similarity_q8 similarity kernel featuring the backward compatibility gate.
  • Added test_blockwise_cosine_similarity unit test verifying mathematical parity ($err < 0.005$ vs $f32$).
• source_rag.rs:
  • Updated HNSW rebuild to bypass embedding_i8 dequantization.
  • Integrated QueryQ8 and cosine_similarity_q8 inside the exact-scan pipeline.
• simple_rag.rs:
  • Updated HNSW rebuild to bypass embedding_i8 dequantization.
  • Integrated QueryQ8 and cosine_similarity_q8 inside simple search paths.
• hybrid_search.rs:
  • Integrated QueryQ8 and cosine_similarity_q8 inside the hybrid exact-scan loop.
• runtime_info.rs:
  • Added dynamic feature mapping for hnsw_streaming_rebuild.

Packaging & Release

• rust_builder/pubspec.yaml: Version bumped to 0.19.2.
• pubspec.yaml (root): Version bumped to 0.20.0 and updated rag_engine_flutter constraint to ^0.19.2.
• CHANGELOGs: Documented the release notes.
• README.md & docs/guides/quick_start.md: Updated to show extended platform support (Windows & Linux).

---

3. Verification & Benchmark Status

Core Unit Tests

All 154 Rust cargo unit tests pass:

cargo test --lib -- --test-threads=1

On-Device Relevance Test (Flutter Drive macOS Profile)

All relevance tests pass:

• default_hybrid Profile: Semantic class passage_recall@10 is restored to 1.000 (100% recall parity).
• vector_only Profile: Semantic class passage_recall@10 is restored to 0.750 (up from degraded 0.625 in the uniform quantization build).