Preserve CP narrow output while planning real batches

Batch-size support needs request-first CP metadata; treating a batch as one long sequence breaks page ownership, top-k ranges, and phase1 compact output collection. This adds a batch CP plan that records per-request page-aligned splits, rank-local offsets, kv/actual-seq metadata, last-token owners, and flattened descriptors for downstream allocator/runtime workstreams.

The scalar full-rerange path now fail-fasts for batch metadata so bs>1 cannot silently discard the narrow-output optimization or restore hidden states with single-request assumptions.

Constraint: CP shared-KV cache state is page-owned and must preserve request boundaries under bs>1.

Rejected: Let bs>1 fall back to scalar full hidden rerange | it loses the phase1 communication reduction and uses wrong single-request metadata.

Rejected: Add a collective to confirm batch plans | all ranks can derive the same plan from CPU metadata and config.

Confidence: medium

Scope-risk: moderate

Directive: Do not remove batch fail-fast guards until W2/W3 consumers use CPSharedKVBatchPlan end-to-end.

Tested: python -m py_compile python/sglang/srt/layers/attention/nsa/utils.py test/registered/unit/layers/test_nsa_cp_utils.py

Tested: remote g0034 container PYTHONPATH=python python -m pytest -q test/registered/unit/layers/test_nsa_cp_utils.py -> 39 passed

Not-tested: full ETE bs>1 CP shared-KV runtime; W2/W3 allocator/direct-write consumers are not implemented yet

docs/advanced_features/nsa_prefill_cp_shared_kv_bs_gt1_implementation_plan_zh.md

@@ -272,8 +272,15 @@ NSAContextParallelMetadata
   request_kv_len_next: List[int]
   request_actual_seq_q_prev: List[int]
   request_actual_seq_q_next: List[int]
+  request_actual_seq_q_prev_cu_tensor: Tensor[batch_size + 1]
+  request_actual_seq_q_next_cu_tensor: Tensor[batch_size + 1]
+  request_rank_local_offsets: List[int]
+  request_last_token_owner: List[int]
+  request_last_token_local_offset: List[int]
   flat_split_list: List[int]
   flat_zigzag_index: List[int]
+  flat_segment_request_ids: List[int]
+  flat_segment_offsets: List[int]
 ```
 
 兼容规则：
@@ -347,9 +354,14 @@ build_batch_page_aligned_in_seq_split_plan(
    - per-request split list；
    - per-request page start/end；
    - per-rank local valid token count；
+   - per-request `kv_len_prev/next` 与 `actual_seq_q_prev/next`；
+   - per-request last-token owner/local offset；
+   - per-request rank-local offset，用于 batch compact hidden collect；
    - flattened segment offsets。
 4. 更新 `prepare_input_dp_with_cp_dsa()`，让 CP shared-KV bs>1 使用新 metadata。
 5. 保留 bs=1 的旧字段兼容。
+6. `cp_collect_last_token_hidden()` 必须继续走 phase1 compact/narrow output；不能因为 bs>1 回到 full hidden gather。
+7. bs>1 full rerange 在 batch-aware full output metadata/kernels 完成前 fail-fast。
 
 测试重点：
 
@@ -371,6 +383,9 @@ cp_size=8
 退出标准：
 
 - metadata 能表达 `batch_size=2`，且不会把两个 request 当成一条长序列。
+- 普通 prefill bs>1 的 output collection 能按 request order 返回 compact last hidden。
+- 需要 full output 语义的 bs>1 请求不会误用 scalar full rerange。
+- 对外 helper 能提供 batch plan、按 plan split tensor、以及 flat page-owner plan。
 
 ### Phase 2：batch-aware owner-lane allocation
 

docs/advanced_features/nsa_prefill_cp_shared_kv_bs_gt1_parallel_workstreams_zh.md

@@ -146,6 +146,14 @@ CPSharedKVBatchPlan
   request_actual_seq_q_prev
   request_actual_seq_q_next
   request_rank_local_tokens
+  request_rank_local_offsets
+  request_actual_seq_q_prev_cu_tensor
+  request_actual_seq_q_next_cu_tensor
+
+  # phase1 narrow output collection
+  request_last_token_owner
+  request_last_token_local_offset
+  output_collect_mode  # narrow_last_token | full_rerange | unsupported_fail_fast
 
   flat_split_list
   flat_zigzag_index
@@ -161,6 +169,8 @@ CPSharedKVBatchPlan
 4. zero-token segment 合法。
 5. batch flattened plan 不能跨 request 合并 segment。
 6. bs=1 继续兼容现有 scalar fields。
+7. phase1 narrow-output 优化不能因为 bs>1 回退：普通 prefill 必须通过 `request_last_token_owner/local_offset` 只收集每个 request 的 last hidden。
+8. bs>1 的 full rerange/logprob/hidden capture 在 batch-aware full rerange 完成前必须 fail-fast，不能使用 scalar `cp_all_gather_rerange_output()`。
 
 ### 对外接口