feat(inference)serve Qwen3-8B directly through Envoy, drop Semantic Router

[sojournerdev]

What

• Removed Semantic Router. Envoy now sends ai.home.arpa straight to vLLM.
• Flattened ai/ into inference/.
• Swapped the model from Qwen3-4B to Qwen3-8B.
• vLLM now deletes the old pod before starting a new one (Recreate), so upgrades can't deadlock on the GPU.

Why

• We're setting Semantic Router aside for now; one less moving part.
• Qwen3-8B is the best model that fits the 8 GB GPU.
• Upgrades used to hang: the new pod waited for the GPU while the old pod still held it.

Checklist

• mise run ci passes