SIGN IN SIGN UP

feat(prefill): always-on per-kind cache-slot census at 32-chunk cadence

The dots3 span-retention hunt burned a day because the by-kind slot
accounting only ran under VMLX_HYBRID_PREFILL_MEM_TRACE, which the
Electron-spawned engine cannot pick up without a full app restart. The
census (slot class x count x resident GB, now with the active reading
inline) runs unconditionally every 32 chunks so the next "active
climbs but nobody knows who holds it" span names its holder from the
ordinary serve log.
J
Jinho Jang committed
2cd0ffc1d0ec2a5d2161cd4473ac13c3d24fb64d
Parent: 5ee0ba6