pi-ts-aperture v0.10.0

46fd304: Dedicated mode now infers the gateway base URL per model from Pi’s native model registry, fixing non-/v1 upstreams like Z.ai in dedicated mode.

Previously dedicated mode registered every OpenAI-compat model at gateway/v1, so Z.ai (/api/coding/paas/v4) produced /v4/v1/chat/completions and 404’d. The extension now cross-references each gateway model against Pi’s registry (by provider id, then model id) to read the upstream base URL, then applies the shared shouldUseGatewayRoot rule: providers whose upstream base URL ends in a non-/v1 version segment (Z.ai /v4) register the gateway root; root baseurls (Mistral, DeepSeek) and /v1 baseurls (OpenAI, Groq) keep gateway/v1. Gateway providers with no native Pi match keep gateway/v1. The inference runs at sync time and the resolved per-model base URLs are baked into the on-disk cache so registerCached replays them before the first revalidation.

Fixes #27 (dedicated mode).