v1.06

Knowledge Base: Streaming Hypergraph Multi-Pass Evaluation (Entry 035)

New community-contributed knowledge base entry from an AAE session optimizing FREIGHT streaming hypergraph partitioning (pybind11 bindings in CHSZLabLib).

Key techniques (1.82x speedup):

• Per-net bit vectors with __builtin_popcountll replacing vector<vector<int64_t>> reverse mapping + std::set-per-net evaluation
• Objective-specific evaluation: direct CUT_NET counting for cut-net mode (no bit vectors needed)
• Incremental bit-setting during the main partitioning loop
• Copy elimination via memcpy, last-pass snapshot skip, pre-allocation

Detailed "what didn't work" section covering loop fusion (hurts ILP), LTO (icache regression), software prefetching (low average degree), code path duplication (icache pressure), and raw pointers vs pybind11 unchecked proxies (pybind11 generated better code).

Full experiment log with 15 iterations (8 kept, 7 discarded).