AutoEP: preserve native Qwen3.5 wrapper text routing

[hazelduan]

AutoEP rejects the Qwen3.5 MoE wrapper and diverges from its native router semantics. This change accepts text input through both the text backbone and the wrapper's language-model path, reads nested text configuration, and preserves native sorted top-k routing, gate-projection score dtype (including BF16 autocast with FP32 inputs), and raw router logits.

It uses the existing preset/spec/router extension points, with defaults preserving other presets. The regression compares Transformers' native models for both classes, with and without CPU BF16 autocast. Vision input remains untested.

Split from #8784; no HiFloat8 or NPU permutation dependency. Related RFC: #8777. Base: 3b80932; initial validated source: c0eb9e1. An open-PR search for Qwen3.5 returned no match on 2026-10-14.

Validation:

• macOS CPU, Python 3.12.2: initial AutoEP suite 194 passed, 6 CUDA-only skips; final four native-router/text-output cases 4 passed. Both autocast cases fail before the gate-dtype correction and pass after it. Changed-file pre-commit and git diff --check passed.
• A5 software environment (Ascend 950PR, driver 25.7.rc1, CANN 9.1, Python 3.11.15, PyTorch/TorchNPU 2.11.0, Transformers 5.16.1): final standalone AutoEP suite 196 passed, 6 CUDA-only skips. The four Qwen parity cases use CPU tensors, not NPU tensors. Command: PYTHONPATH=<isolated-qwen-repo> python -m pytest tests/unit/v1/moe/test_autoep_unit.py -q -rs.

The private combined-stack validation also completed Qwen3.5-35B-A3B text-only EP4/ZeRO2 for 30 matched steps. All 40 layers execute; only routed/shared experts in layers 0/39 train (12 tensors/rank), batch 1/GAS1, sequence length 256, eager attention/experts, no activation checkpointing or checkpoint save, seed 42, AdamW 1e-5/cosine, same model weights and 400-row data. Every rank has exact input hashes, loss, gradient norms, native operator counts and all final trainable tensors. Maximum-rank steady-step median (excluding 5 warm-up steps) is 4.1886s baseline vs 4.1906s candidate (+0.046%); training NPU peak is 33,487,014,400 bytes and initialization peak is 86,776,612,352 bytes in both. This combined-stack result does not attribute an independent speedup or numerical improvement to this split.

Native commands and logs are retained at /data/dxq/hif8_a5_20260914/experiments/quantize_convergence_20261014_01: bash native_gates.sh, bash paired_ep4.sh, qwen_native_stack.log, permutation_native.log, comparison_ep4_baseline_r1_ep4_candidate_r1.json. The runner bind-mounts the original Conda environment (dxq_hif8_torchao_py311) and CANN/driver; it changes no global packages or device permissions. Standalone runtime file hashes match each immutable source revision below. The model run uses a private composition of main PR and split runtime hunks, with the overlapping dispatch hunks reconciled; this composition is not the main PR contents.

Standalone model parity covers BF16 text inference at EP1 with eager attention and sequential experts. Vision, FP16, CUDA-native parity, ZeRO/offload combinations, activation checkpointing, AutoTP, LoRA and multi-node EP remain untested by this split.

AI assistance was used to implement and review this change; the human submitter owns review and validation.

Final published source: d8045b1. Post-push cleanup only removes three duplicate local imports; all four parity cases pass again on macOS and in the A5 software environment. Changed-file pre-commit passes. Runtime source is identical to the initial validated revision.