Struct GptOssLayer 

pub struct GptOssLayer {
    pub router_bias: Vec<f32>,
    pub expert_bias: Vec<ExpertBias>,
}

The per-layer weights the gpt-oss graph carries and the generic GQA layer structs do not.

Held as a side table on Decoder rather than as new Option fields on AttnWeights/MoeWeights for two reasons. The first is mechanical: those two structs have thirty construction sites across seven loaders and every dedicated engine, and none of them will ever set these. The second is the point of the exercise — a checkpoint either has the whole gpt-oss graph or none of it, so Decoder::gpt_oss.is_some() is a single, checkable predicate for “this model needs the gpt-oss path”, which is what the CPU-only and paged-attention refusals below key off. Scattering four independent Options would make “half the graph is wired” representable, and that state is precisely the silent-wrong-answer bug this work exists to remove.

The attention sinks used to be the fifth field here and are AttnWeights::sinks now, because they are NOT gpt-oss-only: mimo2.cpp:58,177 passes the same tensor into the same build_attn_mha. What keeps the whole-graph rule intact is that the loader refuses a gpt-oss file whose sinks are absent.

Fields

router_bias: Vec<f32>

blk.N.ffn_gate_inp.bias, added to the router logits. (attn_output.bias used to be here too and is AttnWeights::o_bias now: crate::proj_bias fills that slot for every graph that creates the tensor, gpt-oss among them.)

expert_bias: Vec<ExpertBias>

blk.N.ffn_{gate,up,down}_exps.bias, one entry per expert.