fix: use max_context_len for stable num_splits in paged decode

- PagedKV::host_kv_len now returns max_context_len instead of max_seq_len
- Eliminates grid-z instability for CUDA graph capture/replay
- Restore skip_no_kernel re-export accidentally removed by ruff --fix

csrc/kernels/attn_kv_source.cuh

@@ -111,7 +111,7 @@ struct PagedKV {
         return p.max_q_len;
     }
     HOST_DEV_FORCEINLINE int host_kv_len(const AttentionParams<bf16>& p) {
-        return p.max_seq_len;
+        return p.max_context_len;
     }
 
     // prefill: Q rows start at qo_indptr[batch] (ragged batch base)

tests/extension/conftest.py

@@ -5,6 +5,7 @@ import torch
 
 from astrai.config.model_config import AutoRegressiveLMConfig
 from astrai.model.transformer import AutoRegressiveLM
+from tests.conftest import skip_no_kernel  # noqa: F401 re-export for test modules
 
 D = 64
 CFG = dict(