v0.4.0 — Artifact Disk Cache

What's New

Artifact disk cache — CLI warm search drops from 14s to ~850ms (16x faster).

All precomputed artifacts (cruncher, holo, spectral, predictive, fingerprints, self-model) are now cached to .graphiq/cache/ as zstd-compressed files. Cache is validated against DB stats on every search and automatically invalidated on reindex.

Cache format details

• HoloIndex: name_holos cached as flat f32 buffer (f64→f32 quantization, ~42MB). Query-term FFT vectors computed on-the-fly at query time (~microseconds for 5-10 terms).
• PredictiveModel: Top-200 per symbol by KL divergence from background, shared vocab index with u32 indices and f32 values (~8.8MB vs 718MB full).
• Cache invalidation: DB stat checksum (symbol/edge/file counts). reindex and upgrade-index explicitly invalidate.

Performance

Metric	v0.3.0	v0.4.0
Cold search (first run)	~29s	~36s
Warm search (cached)	~14s	~850ms
In-process query (MCP)	~18μs	~18μs
Cache size	24MB	~75MB

Search quality

Zero regression — scores identical to 3 decimal places across all 3 benchmark codebases. f32 quantization of normalized vectors is lossless for cosine similarity.

Binaries

• macOS ARM: graphiq-aarch64-apple-darwin, graphiq-mcp-aarch64-apple-darwin, graphiq-bench-aarch64-apple-darwin
• macOS Intel / Linux: pending cross-compilation

Full Changelog: v0.3.0...v0.4.0