distro matrix: stage H could only ever fail, in three more ways

[cnygaard]

The teardown check reported VLLM_GONE:False on ubuntu:26.04 across two releases, held up the 0.8.7 tag, and is written into the 0.8.7 release notes as a possible VRAM leak.

There was no leak. The same run's own log said stopping vLLM — the GPU is free again, and its nvidia-smi reported no compute apps on the card. Only the process bookkeeping disagreed, and the bookkeeping was broken.

Three defects, all in the test

1. pgrep -f 'vllm|VLLM' matched PID 1. In a container PID 1 is the bash -c <script> running this test, and the script's own text contains vllm — --components core,vllm, and the pgrep pattern itself. PID 1 never exits while the script runs, so the "are they all gone?" loop could never succeed:

VLLM_PIDS:1 4475 4562      ← PID 1 is the test's own shell
VLLM_GONE:False            ← foregone conclusion

2. The UI was curled once, immediately after the readiness loop. But that loop waits on vLLM's /v1/models, and glq-chat starts gradio only after the server answers — so there is a window where the API is up and the UI is not. A single curl wins or loses by luck: it passed on 2026-08-17 and failed on 2026-08-18 with identical code. Now polled, with the same die-early check the readiness loop uses.

3. GPU_APPS_AFTER was printed but never asserted — and it is the one signal process bookkeeping cannot fool. An empty compute-apps list is the claim being made, so it is now an assertion.

Verified

RTX PRO 6000 (sm_120), no CUDA Toolkit, glq 0.8.7 installed from source:

EXT_OK:True   GEN_TEXT:Paris. Paris is the most visited city
CHAT_READY:True   UI_OK:True   CHAT_TEXT:Paris. Paris is the most visited city
VLLM_PIDS:4398 4485      ← PID 1 absent for the first time
VLLM_GONE:True
GPU_APPS_AFTER:          ← card empty
stopping vLLM — the GPU is free again

1 passed in 417.62s

Teardown is now confirmed three independent ways: every vLLM PID gone, nvidia-smi clean, and the supervisor's own stop message.

Follow-up

The 0.8.7 release notes carry the retracted claim under "Scope and gaps" — it says the container leg reported vLLM surviving the stop signal. That is now known to be a test artifact and the notes should be corrected; not done in this PR since it edits a published release rather than the tree.

This run also re-confirms, on a second Blackwell box, that 0.8.7 works end to end on sm_120 with no CUDA Toolkit: install, kernels, FULL CUDA-graph capture, a served model, coherent output via both the API and the UI, and a clean shutdown.