#2793 Community benchmark: RTX 3060 12 GB on an AVX-only Xeon E5-2630 v2 (no AVX2/FMA)

Description

## Title
Community benchmark: RTX 3060 12 GB on an AVX-only Xeon E5-2630 v2 (no AVX2/FMA)

## Summary
docs/INSTALL.md lists `STRATA_ISA_FLOOR=avx` ("Older CPUs (experimental)") as not yet tested on a real AVX-only CPU. This is that test: unit tests, speed (cold/warm), NUMA A/B, needles, HumanEval 0-39 / Japanese / tool checks, 65-minute soak and the quant yardstick on a `STRATA_ISA_FLOOR=avx` source build.

## What changed
Results-only submission (no engine changes). Adds `bench/results/2026-10-16-community-rtx-3060-avx-only-e5-2630v2/` (README.md plus the raw logs and summaries).

## Extra Notes
- Hardware: RTX 3060 12 GB behind PCIe Gen2 x4, 2x Xeon E5-2630 v2 (Ivy Bridge, AVX + F16C only), 252 GiB DDR3, Fedora 44
- Strata 296b59d2, engine 0.1.44.1, source build (CUDA 13.3, sm_86, `-DSTRATA_ISA_FLOOR=avx`)
- Model: Flash-Next GSQ-RCO IQ2_XS, context 40,960, KV int8, MTP spec 4, thinking off, network disconnected during the tests

Results:
- Unit tests: all AVX parity tests pass, no SIGILL; CTest 117/119 (two environmental failures, explained in the README)
- Decode: ~17 tok/s bound to NUMA node 1, 24-26 tok/s across both sockets (`ab_engine.py`, 3 rounds)
- Prompt: 178 / 342 / 352 tok/s at 4K / 16K / 32K; TTFT 22.6 / 47.0 / 91.0 s (cold and warm are the same)
- Needles 9/9 at 4K / 16K / 32K
- Same checks as the 2026-10-09 RTX 3060 report: HumanEval 0-39 37/40 base, 35/40 plus (report: 36/40, 35/40), Japanese 11/12, tools 8/8
- Soak 65.7 min, 116 requests, 0 failures, 29/29 known-answer checks
- Quant yardstick: two IQ2_XS runs byte-identical; IQ2_XS vs UD-IQ4_XS top-1 80.3%, KL 0.309

Limitations: one machine, IQ2_XS only for speed, contexts up to 40,960; five mocked setup tests hit the 600 s timeout once in the offline run. On rerun four passed and `test_setup_golden` failed as it does on upstream here (details in the README).

Finding worth noting for others on old dual-socket Xeons: spreading the engine across both sockets was ~40% faster for decode than binding it to the GPU's node.