gfx1151: publish a multi-engine throughput survey for Qwen3.8-27B Q4_K_M, with its correctness status on the face of the table

Row: BACKEND-GATE-ROCM-LLAMACPP

Publish a multi-engine throughput survey for Qwen3.8-27B-Q4_K_M on
strix:gpu0 (gfx1151, Radeon 8060S, ROCm 7.2.4): llama.cpp b10451, the
pinned vLLM 5559679229, and vllm.cpp, on the same board, the same artifact,
the same prompt and the same token count, in one lease, as one table.

Why this exists, and what makes it admissible

The developer directed it on 2026-09-04 — "take numbers from all the engines,
and ours too, so we can compare" — after being told repeatedly that the arm's
declared token gate reads TOKEN_GATE=FAIL and that no deterministic
denominator exists on this path. The authority and its exact scope are recorded
in .agents/developer-preferences.md under "Publish a MULTI-ENGINE SURVEY on
strix:gpu0".

This is a survey, not a parity claim. #2497 already carries one retraction
for quoting a ratio while presenting it as a gated result. The only thing that
distinguishes this from that mistake is that the correctness status is published
with the numbers rather than omitted from them. So the correctness state goes
on the face of the table, not in a footnote a reader can skip:

Every divergence is a near-tie at about 0.125 nats, one bf16 ULP. No
deterministic denominator exists on this path: llama.cpp's greedy decode is not
stable across its own kernel paths (.agents/oracles/llama-cpp.md), and vLLM
disagrees with itself across enforce_eager on both 27B models tried (#2740, and
#2915 on Qwen3.6-27B). Limb 3 of the ratified methodology is unsatisfied and may
be unsatisfiable here (#2864, #2884).

A reader must not be able to take the ratio without also taking the caveat.

Scope

• Artifact. Qwen3.8-27B-Q4_K_M.gguf, sha256
  7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169,
  17,106,775,008 B — the same file the token gate uses, verified on the worker
  before any timing runs.
• llama.cpp b10451. Already measured at 12.233 tok/s, median of 6 legs,
  spread 0.303% (docs/bench-evidence/rocm-strix-llamacpp-denominator-20260902.md).
  That figure is cited, and its open custody link — no compiler ran in that
  lease, the binaries were carried from 2026-09-01 — is carried forward with it.
  The arm is re-taken inside this survey's own lease as well, because two arms
  compared across two leases are two measurements of two boxes.
• vllm.cpp. New. This is what .agents/scripts/rocm-strix-ourarm-staged.sh
  was staged for; its ratification guard's release condition is now met by the
  recorded developer decision.
• Pinned vLLM. New. Recipe committed under
  docs/bench-evidence/oracle-vllm-gfx1151-20260903/. Run in vLLM's production
  configuration: AGENTS.md §Gates forbids --enforce-eager as a denominator,
  so the torch.compile arm is the headline. The correctness run's
  gpu_memory_utilization and max_model_len were chosen for correctness and
  are not inherited blindly.

Methodology

Matched to the landed denominator's, so the numbers are comparable: repeated warm
legs with the leg order recorded and rotated across rounds; N from the design,
never from a log tally; AMD clock windows sampled per leg; the executed kernel
path pinned as bytes, not only the revision; the artifact verified on the worker;
contention state and fleet-at-submission recorded; the arms interleaved rather
than run in blocks, so drift cannot be mistaken for a difference.

Per engine: median, mean, min/max leg, spread as a percentage of the median, and
the underlying repetition range. Then the ratios. Both absolutes always accompany
any ratio. Board faults are reported as a rate, never silently dropped.

Deliverable

One docs/benchmarks/<benchmark-id>.md detail file and its index row in
docs/BENCHMARKS.md, plus the evidence artefacts.

Related: #2497 (which this closes if it succeeds), #2534, #2864.

Read .agents/developer-preferences.md and the existing denominator evidence before using .agents/scripts/rocm-strix-ourarm-staged.sh. Inspect the pinned vLLM recipe under docs/bench-evidence/oracle-vllm-gfx1151-20260903/ and reproduce the specified matched runs on strix:gpu0. Done means a table covering all three engines with correctness status, evidence artefacts, and an index row in docs/BENCHMARKS.md.