#2237 bench: community report, RX 6900 XT (gfx1030) x1 and x2 on 0.1.43.1 - IQ3_S and UD-Q4_K_XL, stock and with the open gfx103x PRs

Beschreibung

## Title
Issue: none. A results-only report; the follow-up of #1927 (the same machine and method on 0.1.42).

## Summary
2x RX 6900 XT 16 GB (gfx1030, PCIe 4.0 x8 each), Ryzen 5 5600X, 128 GB, Ubuntu 26.04, ROCm 10.0. Strata **0.1.43.1**
built from the tag (stock) and with the open gfx103x pull requests (#1149, #1151, #1167, two small local commits, and
#2226), GSQ-RCO IQ3_S and Unsloth UD-Q4_K_XL, one card and the two-card layer split, 4K / 32K / 128K, two runs per
cell, 256 tokens out, greedy. Ten arms, 0 stalls, 0 failed requests.

- **0.1.43.1 reads prompts exactly as 0.1.42 did** (all eight comparable arms within 1% of #1927).
- **Stock decode is unchanged**, except the IQ3_S split with `--pipeline-windows 2`: medians +4% / +8% at 4K / 32K
  (only the 32K ranges do not overlap; 128K equal). The UD-Q4_K_XL split without pipeline windows is identical to
  0.1.42.
- **prs decode**: medians 2-7% above #1927's prs arms in eight of twelve cells, ranges overlapping in most; the bundle
  now also has #2226, so this is not attributed.
- **UD-Q4_K_XL on the split**: `--spec 3 --spec-min-p 0.7 --pipeline-windows 2` instead of `--spec 4 --spec-min-p 0.5`
  is +9-16% decode on stock (50.3 / 53.2 / 51.4 tok/s). With the same configuration as IQ3_S it decodes 28-32% slower
  than IQ3_S, not the 35% #1927 reported with the slower configuration.

| layer split, median tok/s, 4K / 32K / 128K | prompt | decode |
|---|---|---|
| IQ3_S, stock | 462 / 764 / 870 | 73.6 / 78.1 / 71.8 |
| IQ3_S, prs | 975 / 1,782 / 2,033 | 79.4 / 81.6 / 77.7 |
| UD-Q4_K_XL, stock, spec 3 / 0.7 + pipeline windows 2 | 462 / 762 / 869 | 50.3 / 53.2 / 51.4 |
| UD-Q4_K_XL, prs, spec 3 / 0.7 + pipeline windows 2 | 744 / 1,737 / 1,968 | 50.9 / 51.9 / 51.2 |

One-card arms, ranges, TTFT, memory and the 0.1.42 comparison for every cell are in the README.

## What changed
- `bench/results/2026-10-13-community-rx-6900xt-ud-q4_k_xl-iq3_s-0.1.43.1/`: `README.md`, `benchmark.py` (the earlier
  reports' script, unchanged), `table.py`, and per arm `config.json` (local paths rewritten), `summary.json`,
  `results.json`, `BUILD.txt`. No engine or docs change.

## Extra Notes
- Not tested: recall, vision, concurrency, sampling, answer quality of the two formats; two runs per cell.
- Both cards ran with a 245 W board power cap (default 255 W); the earlier reports from this machine did not state it.
- The prs arms bundle several open pull requests and switches; only the bundle was measured.

Developed with an AI coding assistant; every number above was measured on 2x RX 6900 XT (gfx1030, PCIe 4.0 x8 each) / Ryzen 5 5600X, ROCm 10.0.