Goal: measure whether x86-64-v2 (Westmere / SSE4.2 + POPCNT) compilation gives a measurable benefit on the R710 dual-X5675 for the inference and boot paths.

STATUS: PLANNING ONLY. No packages built yet — the running system stays on stock packages so the benchmark baseline is not polluted.

Candidate packages (the "meat"):
- Inference runtime: glibc, libgomp, libm, libstdc++, libgcc_s (what llama.cpp links/executes)
- Model/data path: zstd, zlib (load-time decompression)
- Boot path: kernel-default, systemd

Measurement plan (clean A/B):
1. Snapshot baseline on stock packages: llama-bench q4_0 -t12, systemd-analyze boot time.
2. Build the v2 variants here (RPMs land in this repo, NOT auto-installed).
3. Controlled install + identical re-benchmark + compare.
4. Roll back if gain is within noise.

Expected: small gain (workload is memory-bandwidth bound, not CPU-instruction bound). The goal is to MEASURE it, not assume it.