#3436 [Benchmark] add rx 7900 xtx qwen3.8 flash-next iq3xxs community + standard-v1 results (v0.2.1, engine 0.1.45)

Beschreibung

## Problem and result

No community numbers existed for a single-GPU RX 7900 XTX (gfx1100) Windows
HIP box. This PR adds two result folders measured on 2026-10-20 with
Qwen3.8-Flash-Next GSQ-RCO IQ3_XXS, context 131072, KV int8, MTP spec 4:

- Community report `bench/results/2026-10-20-community-rx-7900-xtx-qwen3-8-flash-next-iq3-xxs/`
  with the legacy suite across four configurations (0.1.43.1 untuned, 0.1.43.1
  with the issue 2274 tuning, 0.1.45 tuned, v0.2.1 serve layer tuned).
- Standard-v1 run `bench/results/2026-10-20-standard-rx-7900-xtx-5b4dbf77/`,
  written by `strata bench run` unchanged, validation included.

Headline numbers: tuned 32K prompts read at 1,818 tok/s, 16K at 1,822 tok/s,
chat decode about 70 tok/s, agent decode 82.1 tok/s. The v0.2.1 serve layer
changes every cell by less than the run-to-run spread against the 2026-10-19
0.1.45 record.

The standard-v1 sanity checks scored 0 of 12: that engine start drew the
per-start bad state of #2960 (the `KV streaming: 100.00% of 0 block reads`
marker on every request). The prompts were provably intact (frozen fixture
token counts on all 22 requests, byte-exact template render), and the
configuration is exonerated by a follow-up replay across four configurations
that answered correctly on five fresh starts. The run stays a valid
performance measurement; the quality result and its diagnosis are recorded in
the README.

## Related discussion

- #2274 - the prompt-path tuning these runs measure (five env flags, prefill
  chunk 16384, hipBLASLt table).
- #2960 - the per-start engine state that produced the 0 of 12 sanity score. A
  comment with this run's evidence follows this PR.
- #3188 - the `strata_bench.py` NameError. A local three-token patch was
  needed to run the legacy suite; a comment with the patch follows.

## Approach

Results only, no engine changes. Raw `results.json` files are copied unchanged
(the three earlier ones under `comparisons/`), the used configuration is
attached, `repeat.sh` holds both commands with their preconditions, and every
number in the README tables was cross-checked against those files after the
commit.

## Validation

- Passed: `strata bench report` on the standard run, status complete, all 89
  missing-map entries carry reasons.
- Passed: 21 of 21 standard requests finished, 0 failed, no unexpected cache
  reuse.
- Failed: sanity 0 of 12 (bad engine start, #2960, see above).
- Not run: the optional concurrency stage, vision (Windows AMD serves
  text-only), the `--model-revision` / `--source-revision` pinning fields.

## Limits

Single machine, one quiet box, warm expert cache (not cold-model numbers), and
the server does not replay seeded requests byte-identically, which limits
exact reproduction. The engine is 0.1.45, not the 0.1.46 this release ships,
because a source install cannot fetch it (`WIN_HIP_ASSET` names an asset no
0.2.x release carries). A separate issue follows for that.

The report text was drafted with AI assistance. Every number comes from the
machine-recorded result files in this PR, none typed by hand or generated.