[shipiyouniao]

Adds KVCached backing for ordinary DeepSeek-V4 C4 KV and CUDA packed FP8/FP4 indexer storage on SGLang 0.5.20. Native full-page IDs drive every selected compressed pool. An allocation miss rolls back new backing leases and native pages before mapping publication; SWA retirement preserves compressed history, while full-page free, grouped free and clear release it.

Depends on #566 and its #570/#400/#565/#376/#539 stack; this adds one C4/indexer commit. Dynamic backing requires TP=1 and PP=1 and managed SWA. Distributed or unknown topology retains native storage/accounting until a cross-rank transaction exists. HiSparse uses separate device IDs and retains native backing. Split AITER FP4, unified/unsupported layouts, compressor state, small pools and shared-draft storage also remain native.

All selected backing managers finish initialization before any resize or reclamation starts. This includes the exact-ID readiness and HiSparse manager-discovery corrections from #570. Regression tests hold each side pool pending and verify that no mapping or resize is published early.

ENABLE_KVCACHED is checked before runner-ownership validation: unset, false, and 0 retain native pools, including direct compressed-pool construction. Native runner tests cover enabled and disabled construction.

Validation at 3c67bf6: 452 targeted CPU tests, GPU-free C++, all-file pre-commit, MyPy 3.10–3.13 and private Codex review passed. A single T4 passed 12 native SGLang storage/runner cases and 4 exact-ID shrink/regrowth cases, with no skips. Tests cover packed byte layouts, reads/writes, rollback and lifetimes using a fixed small budget, without weights, attention/compressor execution or graph capture.

Complete DSV4/MTP serving validation has not been run for this revision. Dynamic TP/PP backing needs asymmetric multi-rank failure/recovery tests. Performance remains unmeasured: compare with native on identical hardware/model/workload for throughput, TTFT, TPOT, tail latency, memory and allocation/reclamation overhead. The T4 pool tests are functional evidence only; no performance target or threshold is claimed.