Build the pinned vLLM oracle for ROCm/gfx1200 in Docker, with a benchmarking/profiling harness against vllm.cpp

Row: ORACLE-VLLM-ROCM-GFX1200-DOCKER

What this is

A standing Docker project that builds the pinned PRIMARY oracle
(e126687a9a, vLLM 0.28.1rc1.dev132) for ROCm on gfx1200, alongside
vllm-gguf-plugin, and runs benchmarking/profiling comparisons against
vllm.cpp's own HIP build on the same box (RX 9060 XT).

This mirrors the measurement BACKEND-GATE-ROCM-LLAMACPP already did by hand
on gfx1151 (#2740,
docs/bench-evidence/oracle-vllm-gfx1151-20260903.md), but packages the
recipe as reusable infrastructure instead of a one-off rc job, and targets
gfx1200 instead of gfx1151.

Why

• .agents/oracles/vllm.md's device-scoped gateability table is empty at the
  CURRENT pin (e126687a9a) on every device. Nobody has built or run the
  primary oracle on ROCm at this pin yet.
• gfx1200 is confirmed in HIP_SUPPORTED_ARCHS and the default
  PYTORCH_ROCM_ARCH at this exact commit (checked directly against
  CMakeLists.txt:52 and docker/Dockerfile.rocm_base:109 at
  e126687a9a), so the prerequisite that blocked earlier attempts on other
  devices does not obviously apply here — unmeasured, not assumed.
• This box's own GPU (RX 9060 XT, gfx1200) is the only local hardware
  available, so a CUDA oracle image is fleet-only; a ROCm image is what
  makes apples-to-apples comparison possible without a lease.

Scope

• Dockerfile building the pinned vLLM commit for
  VLLM_TARGET_DEVICE=rocm PYTORCH_ROCM_ARCH=gfx1200, plus
  vllm-gguf-plugin at its recorded pin, plus the five packages
  oracle-vllm-gfx1151-20260903.md found missing from a bare ROCm image
  (python3-dev, rocm-libs, libdrm-dev, ROCm torchvision, amdsmi).
• A platform-identity smoke test baked into the image (mirrors
  gen_rocm.py's asserts): HSA_OVERRIDE_GFX_VERSION unset,
  RocmPlatform resolved, not UnspecifiedPlatform.
• A benchmarking/profiling script comparing the containerized vLLM oracle
  against a host-built vllm.cpp binary: same prompts, same token counts,
  same GPU (device passthrough), GPU-clock-state discipline per
  .agents/benchmarking.md.

Out of scope

• Advancing the vLLM parity pin or any src//include/ change.
• A performance verdict of any kind before the declared token gate passes
  (per AGENTS.md §Gates); this row builds the tool, it does not itself
  license a benchmark claim.
• CUDA oracle packaging (fleet-only, separate row if wanted).

Blocker to record

This session's user does not have docker group membership on the box the
work is being done on, so the image could not be built or run to verify
inside this session. sudo usermod -aG docker <user> plus a new login
session unblocks it.

Start with .agents/oracles/vllm.md, AGENTS.md §Gates, .agents/benchmarking.md, and docs/bench-evidence/oracle-vllm-gfx1151-20260903.md, then inspect the pinned vLLM commit and existing Dockerfile conventions. The work is done when a gfx1200 ROCm image builds with the specified pins and platform smoke test, and its harness compares it with the host vllm.cpp build under the declared benchmarking discipline.