serving-llms-on-epyc: behavioral tests and supporting checks

[amd-lalithnc]

Split out of #83 so the walkthrough can be reviewed separately. Adds the serving-llms-on-epyc behavioral and unit tests and the code they cover: EPYC 9000-series (Genoa/Turin/Venice) detection, the vLLM 0.25.1 image pin with a runtime stack-compatibility check, and chat-template-based endpoint selection.

[shailensobhee]

Approving. I re-ran the gates locally rather than relying on the PR body; details below.

Verified locally (commit c9bb6da)

• Structural gate: .github/scripts/check.sh -> 0 error(s) across 6 skill(s), plus "Cursor marketplace manifest is up to date" and "Codex plugin and marketplace manifests are up to date". No derived manifest was hand-edited.

• Staleness check: the branch is 7 commits behind main (including lemonade-router-builder from #110), so I applied this PR's content onto current main and ran main's validator against it: 0 error(s) across 7 skill(s). The branch being behind is not breaking anything.

• Unit tests: 19/19 of the pure (non-LLM) tests pass. Detection, the Venice stack-compatibility verdicts, template classification, and endpoint selection all behave as documented.

• Scripts run live: detect.py and validate.py both emit correct JSON. check_model.py was run against the real HF + vLLM registry:

  • Qwen/Qwen3-0.6B -> chat_template.status: present, primary_endpoint: chat_completions
  • allenai/OLMo-2-0425-1B -> chat_template.status: absent, primary_endpoint: completions

  So the registry parsing and the base-vs-chat split work against the real source, not just in the unit tests.

• New image pin exists: amdih/zendnn_zentorch:vllm_v0.25.1_zentorch_v2.11.0.3_ubuntu22.04_2026_ww30 resolves on Docker Hub (digest sha256:027bf7bcdf021ff52211c98514318c7b2dfc95ba9342943367c46c7ca90eb156, pushed 2026-07-29). vllm_version and the image tag were updated together, and test_default_stack_pin_is_consistent enforces that going forward.

• Docs to code: no mismatch found. primary_endpoint and supported_endpoints are genuinely emitted (check_model.py:235), matching what SKILL.md and reference.md tell the agent to read.

On the red checks

The behavioral jobs are failing for CI-infrastructure reasons, not because of this PR. Every one of the 25 errors is the same message:

Failed: claude API not reachable -- are you on the right network?
(... "result":"Not logged in - Please run /login" ...)

That comes from the session-scoped autouse preflight in eval/behavioral/conftest.py, a shared file this PR does not touch. Because the fixture is session-scoped, it errors out the entire module, including the 20 tests that need no LLM at all. The same failure also hits lemonade-router-builder, local-ai-use, and local-ai-app-integration, none of which this PR touches. The runner's claude CLI is simply not logged in.

Worth considering as a separate follow-up: scoping that preflight to only the tests that actually call claude (a fixture on the behavioral tests rather than session-autouse) would let the pure unit tests report honestly when the judge is unavailable.

Non-blocking nit

Two em dashes were added in SKILL.md (lines 352 and 357, the /dev/shm bullet). Not blocking, and consistent with existing prose elsewhere in the repo. Flagging only for house-style consistency.

Nice catch on the --shm-size / --ipc=host conflict, and it is applied consistently across data/epyc.json, SKILL.md, and reference.md.