Before swapping models on a prompt-caching LLM backend, three Jinja2 patterns in the chat template will quietly break your cache hit rate. A 5-minute check against the published tokenizer config catches all three — no GPU, no weights, no inference required.

How we mined 24K analyst-curated training pairs from XSOAR close-notes, dodged a polynomial blow-up, filtered out same-rule near-duplicates, and lifted held-out MRR@10 from 0.598 to 0.846 — a 41% uplift over off-the-shelf bge-reranker-v2-m3.

A debugging story about a rotating billing header that quietly busted the prefix-KV cache, plus a SimpleEngine patch to actually carry KV state across turns. Two fixes together turned 108-second turns into 7-second turns.