docs: split the demo's claims by which lane is serving

The sandbox doc listed real weights and a real tokenizer under 'not claimed',
which stopped being true when --olmo-config was added. The block now separates
what is real in every mode, what is real only with a checkpoint attached, and
what is not claimed either way.

The two additions to that last group are the ones a reader would otherwise
discover for themselves: the model is a base model rather than an
instruction-tuned one, and only one request is served at a time because the KV
cache holds a single sequence.

Author: maincode-dave

docs/demo-sandbox.md

@@ -144,16 +144,26 @@ python3 tools/check_demo_sandbox_static.py    # CPU-only contract gate, in CI
 
 ## What the demo does and does not claim
 
-Real:
+This depends on which lane is serving, so it is worth splitting.
 
-- the OpenAI-shaped route, streaming, and receipt surface
-- on the GPU lane: a real kernel dispatch per request through raw KFD/AQL, a
+Real in every mode:
+
+- the OpenAI-shaped route, streaming, and the receipt surface
+- on the GPU lane, a real kernel dispatch per request through raw KFD/AQL, a
   persistent worker owning the device, admission-time KV block reservations with
   capacity and release accounting, and scheduler-visible KV ownership checks
 
-Not claimed:
+Real only with `--olmo-config`:
+
+- real weights, a real tokenizer, and text the model actually generated
+- `synthetic: false` on `/v1/models` and `/api/health`
+
+Not claimed in any mode:
 
-- real weights, a real tokenizer, or chat quality of any kind
+- chat quality. `OLMo-2-0425-1B` is a base model, so it completes rather than
+  converses, and no instruction tuning is involved here
+- concurrency. One request is served at a time because the KV cache holds one
+  sequence, so a second caller waits
 - multi-user scheduling, auth, quota, or abuse controls
 - any throughput number derived from this page