Three improvements for DeepSeek API cost/performance:
1. Daemon cache warming: export DEEPSEEK_API_KEY, COLIBRI_CACHE_WARMING=true,
COLIBRI_CACHE_WARMING_INTERVAL_HOURS=6 in rc.d prestart. Keeps DeepSeek
prefix cache warm across sessions (~3,500 free tokens/request).
2. pi HTTP/2: change allowH2:false → true in http-dispatcher.js after
npm install. Enables HPACK header compression and multiplexing for
API calls (Node 24 + undici 8.3 handle this correctly).
3. Live verification: cache warming triggered and confirmed cache HIT
with 3,584 hit tokens on first warm cycle.
Tested: daemon restart, cache warming log, CLI status confirms enabled.
2605465e7e