The online-serving gate aborts on current main: harness looks for examples/server, the build produces examples/vllm-server

[localai-bot]

The canonical online-serving gate cannot run against current main. It aborts
before starting a server.

examples/CMakeLists.txt declares the CMake target server but sets its
output name to vllm-server:

69:  add_executable(server server/main.cpp)
86:  set_target_properties(server PROPERTIES OUTPUT_NAME vllm-server)

The harness checks the on-disk path, which no longer exists:

scripts/dgx-online-serving.sh:293  [[ -x ${build_dir}/examples/server ]] || {
scripts/dgx-online-serving.sh:294    echo "provenance-recorded build did not produce examples/server"
scripts/dgx-online-serving.sh:382,424,693   same path
tools/bench/online_gate.py:3596    "server": build_dir / "examples" / "server"
tools/bench/online_gate.py:3756    server_path = build_dir / "examples" / "server"

Verified on a clean sm_121a Release build of bc6e3d72: examples/ contains
vllm-server, vllm-bench, vllm-cli, and no server.

Introduced by 1a02ab4f feat(release): package canonical server (W6), which set
the release output name without updating the harness that consumes it.

Why this matters beyond a broken path

--execute is how a 27B/35B online-serving ratio becomes an accepted
measurement. While it aborts, throughput numbers for the gate models can only
come from ad-hoc harnesses that do not enforce the checks this one does — the
pinned revision, the oracle version and dependency inventory, cache-drop proof,
one-model/one-lock, and the recorded run recipe. That is the difference between
a number and evidence.

Fix

Point the harness at the built artifact and add a guard so the two cannot drift
again: assert the path the harness expects equals the OUTPUT_NAME declared in
examples/CMakeLists.txt, rather than hardcoding either spelling in two places.

A red test must fail on today's tree before the fix.

Related

• Qwen3.6 NVFP4 baselines (27B and 35B-A3B) must reach vLLM speed parity #213 (Qwen3.6 NVFP4 speed parity) is blocked on this for any binding number.
• tools/bench/online_gate.py has model keys 27 -> unsloth/Qwen3.6-27B-NVFP4
  and 35 -> nvidia/Qwen3.6-35B-A3B-NVFP4. There is no key for
  nvidia/Qwen3.6-27B-NVFP4, which is the 27B the campaign targets, so that
  model has no reproducible canonical recipe at all. Worth adding in the same
  area of the code.