docs: make the hardware not-claimed list unambiguous about serving

'Production serving through the model API' was technically accurate, since the
OLMo path does not go through the model API, but it reads as a blanket denial
that real weights are served at all. It now separates the two: real weights are
served over an OpenAI-shaped endpoint one request at a time, and the model API
remains CPU-only metadata that the OLMo path does not touch.

Author: maincode-dave

docs/hardware-support.md

@@ -165,8 +165,12 @@ wrapper with a ROCm image that contains `rccl-tests`.
 - Support for AMD GPUs other than the current MI355X/gfx950 target.
 - Support for virtualized or container environments that do not expose `/dev/kfd`
   and `/dev/dri`.
-- Production serving, continuous batching, or OpenAI-compatible API behavior
-  through the model API.
+- Production serving. The OLMo 2 path does serve real weights over an
+  OpenAI-shaped endpoint, but one request at a time, with no batching, no
+  scheduling across sequences, no auth and no quota.
+- Continuous batching, or any OpenAI-compatible API behaviour through the model
+  API. The model API is CPU-only metadata and preflight, and the OLMo path does
+  not go through it.
 
 ## Reporting Hardware Results