fix: respect packed document boundaries in Qwen3.5 DeltaNet CP

[hallerite]

Fixes #3002.

Under Ulysses CP the DeltaNet layer built its fla CP context from a fabricated single-sequence cu_seqlens ([0, global_seq_len]), so the gated delta rule carried recurrent state across packed documents. Independently, the causal convolution derived its seq_idx from shard-local cu_seqlens, resetting mid-document at every CP shard boundary — wrong even for unpacked rows.

The fix hands fla the information it already knows how to use:

• _build_cp_context now uses the full pre-shard cu_seqlens published by setup_cp_params (same source as the Nemotron-H fix in fix: reset Nemotron-H CP state at packed boundaries #3001). fla derives per-rank boundaries, document-aware state passing, and conv halo sizes from them.
• Under CP the convolution runs through fla's CP conv (fla.modules.conv.causal_conv1d with cp_context), which resets at true document boundaries and exchanges tail tokens with the previous rank for documents straddling the shard boundary, in both forward and backward.

The non-CP paths are untouched.

Validation

2-GPU (CP=2, bf16, real fla kernels, layer wired exactly as the trainer wires it), packed row [37, 51, 40] with doc 2 straddling the shard boundary, vs the trusted non-CP varlen path:

check	before	after
doc 2 max|CP − ref| (packed)	4.6e-01	7.8e-03 (bf16 noise)
single unpacked doc, shard-boundary tokens	4.6e-01	2.0e-03 (1 ulp)
doc 3 response to doc-2 tail perturbation, slow decay (must be 0)	7.8e-01	0.0

tests/unit/train/models/test_qwen3_5_moe.py passes (4 passed); ruff check/format clean.

🤖 Generated with Claude Code

---

Note

Medium Risk
Targets distributed-training correctness for linear-attention layers under Ulysses CP and packed batches; scope is limited to the DeltaNet CP path with non-CP behavior preserved.

Overview
Fixes incorrect recurrent state and convolution boundaries in Qwen3_5MoeGatedDeltaNet when Ulysses context parallelism is enabled with packed sequences.

_build_cp_context now feeds fla the full pre-shard cu_seqlens from ULYSSES_PARAMS (same pattern as Ulysses flash attention) instead of synthesizing [0, global_seq_len], so the gated delta rule no longer carries state across packed documents.

When CP is active, the causal conv path switches to fla’s CP causal_conv1d with the same cp_context, so resets happen at real document boundaries and halo tokens are exchanged for documents that span shard edges. Non-CP conv and attention fallbacks are unchanged.

^{Reviewed by Cursor Bugbot for commit 6f068ea. Bugbot is set up for automated code reviews on this repo. Configure here.}