Prevent batched CP draft from silently leaving the target path

EAGLE draft shared-KV is supposed to mirror the target CP layout, so bs>1 must not fall back to legacy full-input or padded-hidden behavior when required batch metadata is missing or inconsistent. This change keeps the existing bs=1 compatibility path but makes batched CP draft fail fast on missing/mismatched spec hidden states, embedding pad metadata, or input embed shapes. The docs record the current W7 boundary: draft prefill follows target metadata, while scheduler admission and ETE remain gated.

Constraint: CP draft KV must mirror target layout and must not silently diverge under bs>1 shared-KV.

Rejected: Allow bs>1 to use the old full-input fallback | it can hide wrong owner/page metadata and corrupt accept length.

Confidence: medium

Scope-risk: moderate

Directive: Do not open the scheduler bs>1 CP gate until EAGLE accept length/output length are verified with this fail-fast path enabled.

Tested: Remote g0034 targeted EAGLE fail-fast unit test passed; remote full test/registered/unit/layers/test_nsa_cp_utils.py passed 70 tests.

Not-tested: EAGLE bs>1 ETE, because scheduler CP bs>1 admission gate remains closed.

test/registered/unit/layers/test_nsa_cp_utils.py

@@ -1,4 +1,5 @@
 import ast
+import os
 from pathlib import Path
 import unittest
 import sys
@@ -33,6 +34,7 @@ from sglang.srt.layers.attention.nsa.utils import (
     split_in_seq_cp_local_pair,
 )
 from sglang.srt.mem_cache.cp_shared_kv_layout import CpSharedKVLayout
+from sglang.srt.models.deepseek_nextn import DeepseekModelNextN
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=1, suite="stage-a-test-cpu")
@@ -714,6 +716,42 @@ class TestNSAInSeqCPUtils(unittest.TestCase):
         ):
             cp_collect_last_token_hidden(torch.zeros((8, 1)), forward_batch, 2)
 
+    def test_deepseek_nextn_cp_draft_bs_gt1_fails_fast_on_hidden_shape_fallback(
+        self,
+    ):
+        import torch
+
+        model = DeepseekModelNextN.__new__(DeepseekModelNextN)
+        model._debug_cp_draft_shared_kv = lambda _message: None
+        forward_batch = SimpleNamespace(
+            uses_cp_shared_kv=True,
+            batch_size=2,
+            extend_seq_lens_cpu=[4, 9],
+            nsa_cp_metadata=NSAContextParallelMetadata(batch_size=2),
+        )
+
+        with (
+            patch.dict(os.environ, {"SGLANG_CP_DRAFT_SHARED_KV": "1"}),
+            patch(
+                "sglang.srt.models.deepseek_nextn.get_attention_cp_rank",
+                return_value=0,
+            ),
+            patch(
+                "sglang.srt.models.deepseek_nextn.get_attention_cp_size",
+                return_value=8,
+            ),
+            self.assertRaisesRegex(
+                RuntimeError,
+                r"\[CP_SHARED_KV_FAIL_FAST\]\[draft_batch_gt1_spec_hidden_shape_mismatch\]",
+            ),
+        ):
+            model._get_cp_local_spec_hidden_states(
+                forward_batch,
+                torch.zeros((3, 2)),
+                full_num_tokens=13,
+                local_num_tokens=8,
+            )
+
     def test_full_rerange_fails_fast_for_batch_metadata(self):
         import torch