CP HiCache write-through keeps a host copy, but current disaggregation registers and addresses prefill GPU KV pages. The document records why CPU-source transfer needs a distinct source contract, host backup readiness, host-slot pinning, and CP logical-page mapping before it can safely release prefill GPU L1 pages early.
Constraint: Current Mooncake KVArgs register GPU token_to_kv_pool buffers as the prefill transfer source
Rejected: Reuse device page_indices as host page ids | host/L2 allocator uses separate physical slots and would corrupt decode destinations
Confidence: medium
Scope-risk: narrow
Directive: Do not switch prefill-to-decode transfer to host source without explicit backup-ack and host-slot lifetime contracts
Tested: Documentation-only change; inspected current prefill/decode/Mooncake/HiCache paths
Not-tested: CPU-source transfer implementation
(cherry picked from commit 01d08e3526ad95c0ab924e6ba859139615fc65c9)
da8bd7f813