dots3-note vision tower: use_bias=true still refuses, because the shared MlpGateUpMethodBase seam has no bias arm

Row: MODEL-MM-dots3-note-dots3-note-for-causal-lm

W6b (#2613) lifted four of the five config arms W6a deferred to it —
adapter_type = pixel_shuffle_mlp, post_norm = false, use_qk_norm = false
and is_causal = true all compute now. use_bias = true is the fifth and it
still refuses BY NAME. This issue owns it.

What use_bias turns on upstream

DotsMoEVitConfig.use_bias (vision.py:43 @ 9035151d6) is threaded into
three places:

• nn.Linear(dim, dim*3, bias=params.bias) and nn.Linear(dim, dim, bias=params.bias) — the attention qkv and proj
  (vision_attention.py:143-144).
• DotsSwiGLUFFN(..., bias=bias) — fc1, fc2 and fc3 of every DENSE block
  (vision.py:129-134).
• DotsSwiGLUFFN(self.hidden_size, config.moe_intermediate_size, bias=config.use_bias) — every ROUTED EXPERT's fc1/fc2/fc3
  (vision.py:158-163).

So a use_bias = true checkpoint carries 5 extra tensors per dense block and
2 + 3 * num_routed per routed block. On the released geometry that would be
1949 tensors nothing in the tree reads.

Why W6b did not lift it

1. Nothing ships it. dots-studio/dots3-note-prev sets use_bias: false,
   and DotsMoEVitConfig's own default is False (vision.py:43). No
   published dots3-note checkpoint sets it true.
2. The shared MLP seam has no bias arm. Both the dense block and every
   expert reach their SwiGLU through layers::MlpGateUpMethodBase
   (include/vllm/model_executor/layers/linear.h:82), whose Apply returns
   silu(gate) * up from weights alone. Lifting use_bias means either adding
   an optional bias to that base — an edit to the seam every model in the tree
   routes its MLP through, for a configuration none of them has — or writing the
   two GEMMs by hand beside it, which is the parallel path AGENTS.md's "Shared
   seams" forbids.
3. It would land unreached. AGENTS.md's "Nothing lands dead" asks what
   production entry point reaches new code at its own merge commit. A
   use_bias arm is reachable only from a checkpoint that declares it, and the
   only such checkpoint would be a synthetic fixture written to reach it.

Refusing by name is therefore the honest state: the message says which keys the
checkpoint set and what is missing, and the text tower of such a checkpoint
still loads beside the refusal.

Closing it

Add an optional bias operand to layers::MlpGateUpMethodBase and its
unquantized members (red-first, with the existing MLP gates proving the
no-bias path is byte-identical), load the qkv/proj/fc biases, and delete the
refusal in the same change. Gate it the way W6b gates the other lifted arms: a
fixture whose use_bias is true, served end to end through
ApiServer::handle_chat_completions, against the in-test double-precision
reference.

Start with AGENTS.md, layers::MlpGateUpMethodBase in include/vllm/model_executor/layers/linear.h, and the bias wiring in vision.py and vision_attention.py. Trace W6b's fixture through ApiServer::handle_chat_completions and compare it with the in-test double-precision reference. Done means a true-use_bias fixture loads qkv, projection, dense-MLP, and routed-expert biases while preserving the existing no-bias path.