feat(search): exact residual cascade CPU core (chain 2/3)

[christopherkarani]

Chain Info

Chain: 2 of 3
Strategy: feature-branch-chain
Base: chain/1-extract-selection (merge 1 first)

• ✅ PR refactor(search): extract flat-search top-k selection (chain 1/3) #28 — slice 1: selection extraction refactor
• 🔄 THIS PR — slice 2: exact residual cascade core
• ⬜ PR 3/3 — slice 3: Metal residual bound kernel + GPU routing

Summary

• Adds the exact residual-bound cascade: fp32 PCA projection head built in double precision (cblas_dsyrk + deterministic Jacobi), Pythagorean tail-norm lower bounds, parallel seed selection, exact survivor rescoring, and top-k resolution.
• VectorBuffer gains the cache-invalidation hook the cascade requires.
• No GPU routing yet: this slice is CPU-only and compiles green standalone.

Changes

File	Change
Sources/MetalANNSCore/ResidualCascadeMath.swift	Double-precision PCA, Jacobi eigendecomposition, query bounds
Sources/MetalANNSCore/ResidualCascadeSelection.swift	Exact rescoring + top-k resolution
Sources/MetalANNSCore/ResidualCascade.swift	Cascade orchestration, eligibility, lazy GPU-plane cache scaffolding
Sources/MetalANNSCore/VectorBuffer.swift	Cache invalidation hook (+3 lines)
Tests/MetalANNSTests/ResidualCascadeTests.swift	Brute-force equivalence across metrics/dims/seeds, invalidation tests

Lines

Diff: +1774 / −0 (core feature slice; split per chain plan)

Test Plan

• Release build clean on this branch alone
• Full suite: 325 tests / 71 suites pass
• Residual cascade is brute-force exact across cosine/L2/inner-product, multiple dims/seeds

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.