…he hard-exclusive one

The pre-router README claimed llama-cpp and Ollama can't coexist on the
GB10 — true for classic single-model mode (-ngl 999 pins all layers at
boot), false for router mode. Router mode is lazy: no VRAM is claimed
until a model is loaded. Ollama is also lazy (loads on chat request,
unloads after keep_alive). So router mode + Ollama coexist as long as
the simultaneous resident set fits the 124 GiB GB10.

vLLM is the actually-exclusive engine in this stack: --gpu-memory-
utilization 0.9 reserves ~90% of VRAM at startup whether or not it's
serving requests. classic single-model mode also remains exclusive.

Replaces the "GPU exclusivity" subsection in llama-cpp/README.md with
"GPU sharing" — a four-row table comparing posture (lazy vs eager) and
a coexistence rule of thumb. Adds an explicit note that Ollama has
`restart: unless-stopped`, so a Docker daemon restart will re-up it
even after a stop; the explicit stop-and-start workflow is still
needed if you want hard exclusivity (e.g. loading the 120b on both
engines simultaneously).

Spec doc gets a corresponding "GPU exclusivity reclassified" note in
the post-deployment-findings section.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>