#2032 community bench: 2x RTX 2080 Ti (sm_75, NVLink unused), IQ3_S, 0.1.43

Description

community bench: 2x RTX 2080 Ti (sm_75, NVLink unused), IQ3_S, 0.1.43

Speed-only report from `tools/strata_bench.py --long` (3 runs per test,
medians with ranges), measured on the machine's own HTTP API.

- Hardware: 2x RTX 2080 Ti 22 GB (44 GiB VRAM), i5-11500T 6C/12T,
  62.6 GiB DDR4-2933, Ubuntu (kernel 7.0.0-28), driver 610.57.04
- Engine: 0.1.43 source build (commit 7695859a), sm_75
- Model: original Flash-Next IQ3_S at the native 262144-token context,
  layer split 24 across the two cards, `--kv int8`, `--prefill auto`
- Results: chat decode 78.9 tok/s; fresh prompts 873.7 / 1521.9 / 1776.9
  tok/s at 4K / 16K / 32K; agent session 68.8% cache hit

The 32K prompt number is about 15% above the 2x TITAN RTX (also sm_75)
report at 0.1.40.3 on the same model and context.

Two notes from this machine that may interest others:

- `STRATA_BF16_TC=1` and `STRATA_PREFILL_HELP=1` were both set. The first
  is a no-op here (7.x cards take the FP16 tensor-core path by default);
  the second measured +18% on short prompts in our own A/B.
- The conversation cache budget only matters if the admission rule can be
  met: with a full expert arena in RAM, MemAvailable after load was about
  2.58 GB, below the "snapshot + 2560 MiB floor" every time, so a large
  `--conversation-cache-mib` produced 166 rejected parks, 21 evictions and
  a first-token spike to 72 s. Lowering the budget to 2048 MiB removed
  both the rejections and the spike and left prompt throughput unchanged.

No correctness check was run; background load and power limits were not
recorded (the cards were on their default policy during the runs).