pi-neuralwatt v0.15.3

51f9dbf: fix: exclude embeddings model from catalog and harden resolveMaxTokens
d15f9c8: Make the model-store cache scaffold-aware. A fresh entry persisted without a key only carries the public scope, so logged-in users never saw preview, grant-gated, or private models — and a key-scoped entry could be replayed for an anonymous user. Each refresh now stamps the persisted entry with the scope it was fetched under (public / key v1) and treats a scope mismatch as stale: keyed refreshes bypass anonymous entries, anonymous refreshes and cache-only boots no longer replay key-scoped entries, and fresh same-scope entries still skip the network. Legacy entries without a stamp re-fetch once and are restamped. Supersedes the behavior proposed in #90.
76b4464: Remove the 327,680 context-window override from the Kimi K3 family. The endpoint no longer rejects requests above 327,680 tokens; both /v1/models (max_model_len, limits.max_context_length) and the live endpoints now agree on 1,048,560 for kimi-k3, kimi-k3-fast, and kimi-k3-flex. The fallback table and drift check follow the API again.