Under the Model Context Protocol (MCP), agentic large language models (LLMs) re‑encode verbose tool schemas at every turn, making the prefill cost quadratic in sequence length and dominating time‑to‑first‑token (TTFT) as the tool registry grows.\ \ Nexus’s primary lever decouples routing from the schema‑prefill cost: an INT8 semantic lookaside buffer (SLB) equipped with a calibrated cross‑encoder margin gate retrieves the appropriate tool, and arguments are generated over a compressed textual signature (median 19 tokens) instead of spliced key/value (KV) cache. This path is depth‑independent; routing accuracy remains around 89% even when the registry scales to 250 tools, whereas a concatenate‑all‑schemas baseline would overflow the context window entirely.\ \ In terms of speed, Nexus delivers the first‑argument token 1.66× sooner and saves roughly 80% of main‑context tokens.\ \ As a secondary, bounded lever, Nexus transplants a compiled schema KV block directly into the live context. This approach is fundamentally limited by rotary position embedding (RoPE) phase drift: an anchored splice is output‑exact, but off‑anchor placement corrupts attention beyond a threshold of $P=256$. Beyond this point, Nexus repairs the seam with a depth‑adaptive suffix redecode that may escalate to a full re‑prefill. The resulting guarantee is on output fidelity (top‑1 agreement, $D_{KL}\approx0$), not on latency, which can dip to about 0.98× before converging to parity. At moderate depth, TTFT speeds up 1.1‑1.7×, narrowing to parity as context depth increases.\ \ Two negative results bound the design: the RoPE fidelity boundary, and the failure of a reference‑free drift gate to predict drift (Spearman $\rho=0.193$).\ \ All measurements are taken from a single model tuple (Qwen2.5‑14B‑Instruct Q4_K_M) on Apple‑silicon unified memory; while the quantitative envelope is tuple‑specific, the qualitative boundaries appear to generalize.\ \ Blogger's Review: Nexus cleverly sidesteps the quadratic prefill cost by leveraging semantic retrieval and depth‑adaptive splicing, making it highly practical for large‑scale tool registries. Although RoPE drift remains a limitation, the "never‑regress" fidelity guarantee provides a solid safety net for real‑world deployments.