[yueming-yuan]

Problem

The R3 routed-experts capture runs at the top of _post_process_topk_ids, ahead
of every padded-region mask in that function. Rows at or past
num_token_non_padded therefore reach the capturer holding whatever the router
left in them, and the capturer writes buffer[:batch] without clearing, so those
rows carry a stale routing row from an earlier forward.

Replayed into training, a padded region comes back as one repeated row of real
expert ids. The R3 replay check counts each as a zero-overlap mismatch.

Evidence

From a miles CI run (DeepSeek-V3.2-5layer, 8192-token batch,
chunked_prefill_size=4096):

• one replay row repeated 67 times across contiguous tokens 4039-4053+, while
  the recomputed routing varied per token
• 104 mismatches against a threshold of 82
• the block starts at the chunked-prefill boundary

Same check on the pre-bump build: 171 replay rows, all 171 unique, zero
repeats, and at most 9 mismatches per check.

Why not just move the capture after the mask

The masks are per-branch and disagree on fill value: the CUDA path fills -1, the
ROCm path fills 0 -- a valid expert id that replay cannot tell from a real
selection. Masking a copy at the capture site gives replay -1 on both, and leaves
the tensor the MoE kernels consume untouched.

The NPU call site is left alone: _mask_topk_ids_padded_region is a no-op there,
so passing the count would only add a clone.

---

CI States

Latest PR Test (Base): ❌ Run #32918538859
Latest PR Test (Extra): ❌ Run #32918538639
Latest PR Test (AMD ROCm 7.2): ❌ Run #32918538816