Route CP shared MLA store through TAI fused kernels without runtime spam

The shared-KV prefill path now optionally calls tai_kernel.nsa_prefill.fused_store_mla_kv before falling back to logical_locs_to_physical plus set_mla_kv_buffer. The fast path supports packed FP8 and BF16/FP16 direct KV buffers, while debug mode and kernel failures still preserve the existing fallback behavior. Success logging was removed after path verification because per-layer/per-rank logs are too noisy in normal server runs.

Constraint: Runtime must remain safe when tai-kernel is absent or debug checks are enabled
Rejected: Keep success logs permanently | floods prefill logs once every rank/layer starts using the fast path
Confidence: high
Scope-risk: moderate
Directive: Keep fallback warnings; do not re-add per-layer success logs outside explicit debug instrumentation
Tested: g0034 container python -m py_compile python/sglang/srt/layers/attention/nsa/cp_shared_kv_runtime.py
Tested: g0034 container PYTHONPATH=python pytest -q test/registered/unit/mem_cache/test_cp_shared_kv_runtime.py -q (40 passed)
Not-tested: Full multi-node PD server throughput after log removal

docs/advanced_features/nsa_prefill_cp_phase7_materialize_triton_plan.md

@@ -474,6 +474,50 @@ SGLANG_CP_SHARED_KV_USE_TAI_MATERIALIZE=1 fast path
 - fallback reason 可解释；
 - profiler 中 `materialize.token.local_copy` / `materialize.paged.local_copy` 下降。
 
+### P0 fused MLA persistent store 接入
+
+在 materialize 之前，shared KV direct-write 的 MLA persistent KV 写入还有一段
+独立开销：
+
+```text
+logical_locs_to_physical(...)
+quantize_k_cache_separate(...)
+set_mla_kv_buffer_triton(...)
+```
+
+对应的 tai-kernel fast path 已通过 SGLang 实验开关接入：
+
+```text
+SGLANG_CP_SHARED_KV_FUSED_MLA_STORE=0/1
+```
+
+默认关闭。开启后只影响 prefill CP shared-KV 的 MLA local persistent write：
+
+```text
+forward_mla.py
+  _maybe_write_cp_shared_local_mla_kv(...)
+    -> try_tai_fused_mla_store(...)
+       -> tai_kernel.nsa_prefill.fused_store_mla_kv(
+              k_nope, k_rope, raw_kv_buffer, logical_locs,
+              page_size, cp_size)
+    -> fallback logical_locs_to_physical + token_to_kv_pool.set_mla_kv_buffer
+```
+
+接入约束：
+
+1. `fused_store_mla_kv` 根据 raw KV buffer dtype 分发：
+   - `uint8 [capacity, 1, 656]`：packed NSA FP8 路径，
+     fused logical->physical + FP8 quant + scale/rope store；
+   - `bf16/fp16 [capacity, 1, 576]`：non-FP8 direct-store 路径，
+     fused logical->physical + MLA KV direct write。
+2. tai-kernel import 失败、shape/dtype 不支持、kernel runtime 失败时自动 fallback。
+3. `SGLANG_DEBUG_CP_SHARED_KV=1` 时强制保留原 PyTorch path，便于继续暴露
+   shared KV correctness 问题。
+4. fused kernel 消费 shared logical token locs；调用前仍要求 direct-write gate
+   已确认 `local_out_cache_loc` 属于当前 CP owner。
+5. 该路径不替代后续 attention 前的 shared KV materialize，只减少 persistent
+   write 阶段的 remap/quant/store kernel 碎片。
+
 ## Benchmark plan
 
 新增 benchmark 只测本地 materialize，不包含 CP all-reduce：