Record model decision (Qwen3-32B FP8) and vLLM SM120 caveat

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

TODO.md

@@ -26,7 +26,8 @@ Server is delivered; base setup + NVIDIA driver are done (see `SETUP.md`).
 - [ ] Role `llm_stack` — docker compose: vLLM (fixed `--gpu-memory-utilization`),
       embedding server, Open WebUI + pgvector, reverse proxy, oikb timer.
       Tag compose tasks with `compose` so `just compose` works.
-- [ ] Pin versions at install time (vLLM / Open WebUI / oikb / model) per §5
+- [ ] Smoke-test a vLLM container on the server (official images may not support SM120)
+      — pin the working image, then build `llm_stack`. Start model: Qwen3-32B FP8
 - [ ] OCR pipeline for ~166k scanned PDFs (§2.1) — own work block, competes with vLLM for the GPU
 - [ ] TLS: currently plain HTTP on `chat.phytron.local`; retrofit an internal CA
       certificate (AD passwords travel in clear text until then)