#2439 docs: community benchmark — RTX 5080 + 4070 Ti unequal pair, both orders (+ helper cache arm)

本文

## Summary
- Adds `bench/results/2026-10-14-community-rtx5080-rtx4070ti/` answering the unequal dual-NVIDIA ask in `docs/TEST_REQUESTS.md`: **both card orders** plus a helper-expert-cache arm and a single-card control on v0.1.42.
- Medians of 3 runs per prompt tier (8K/70K/240K real tokens, per-run cache-busting salt), 5-run decode arm; per-run JSONL in `data/`.
- Headline: single 5080 won every tier (decode 85.0 tok/s); both split orders lost 30-55% prefill / ~12% decode; helper cache (`--expert-cache-device1 5000 --remote-expert-opt`) collapsed decode (23.6, spread 16.6-48.5).
- Records one crash condition as a repro datum, not a number: layer split + 240K KV streaming hard-crashed the machine (dump write failed) — issue #2438.
- Also documents the P2P probe (`cuDeviceCanAccessPeer`=0 both ways on WDDM GeForce).

Related: #2438 (crash repro), #2052 (asymmetric-pair defaults)