fix(MODEL-TEXT-commandr-cohere-for-causal-lm): a Cohere config carrying rms_norm_eps runs mean-centred LayerNorm silently

Row: MODEL-TEXT-commandr-cohere-for-causal-lm (.agents/model-matrix.md:216, BLOCKED)

Found by wave PORTQ-5 (#2679) re-deriving PORT-NOW entry [190], upstream aa9903490c vllm#50156. Nothing was executed.

One of the commit's four halves is portable

Upstream adds an RMSNorm class and select_norm_impl(config), which picks RMSNorm + rms_norm_eps when that key is present and LayerNorm + layer_norm_eps otherwise, applied to q_norm / k_norm / input_layernorm / the final norm.

This tree hard-wires Cohere LayerNorm with no selection and no refusal:

• src/vllm/model_executor/models/commandr.cpp:52-57 — LayerNormInto(...) -> vt::LayerNorm(..., /*bias=*/nullptr, ...), called at :159 (input_layernorm) and :209 (final norm).
• Eps comes from one place: src/vllm/model_executor/models/commandr_registry.cpp:108-111, RawDouble(config.raw, "layer_norm_eps", 1e-5).

Nothing reads rms_norm_eps (zero hits across the four commandr files; positive control through the identical probe form: layer_norm_eps => commandr_registry.cpp:108,110 and commandr.h:98). So a checkpoint of the new Cohere flavour would run mean-centred LayerNorm at a 1e-5 default in place of RMSNorm — a silent numerics divergence, not a refusal, which is the part that makes this worth filing against a BLOCKED row.

The other three halves are not portable

• head_dim: already satisfied, at the shared config layer rather than in the model — src/vllm/transformers_utils/hf_config.cpp:448-452 honours an explicit head_dim > 0 and derives otherwise, consumed at commandr.cpp:90, commandr_weights.cpp:60, commandr_registry.cpp:119,143.
• The sliding_window + 1 training convention: surface-absent. Sliding window is refused on this arch at parse (commandr_registry.cpp:134-138), matching upstream's own if not self.v1 guard. The cohere2_moe.py half has no file here; that belongs to MODEL-TEXT-cohere2-moe-cohere2-moe-for-causal-lm (.agents/model-matrix.md:217, INVENTORIED).
• The single fused all-reduce of attention + MLP: surface-absent. commandr here is single-device — no TP symbol occurs in either commandr TU or its header, and TpAllReduceSum exists in exactly one model TU (src/vllm/model_executor/models/qwen3.cpp:148). With no two all-reduces there is no pair to fuse.

Size, and an honest cheaper option

Full fix: an RmsNormInto sibling to LayerNormInto (vt::RmsNorm already exists), a CommandrNormKind / eps resolver beside CommandrLayerNormEps, threaded through commandr.cpp:159,209. The q_norm / k_norm arm is moot because use_qk_norm is refused at parse. ~40-60 lines product + ~40 lines test.

Cheaper and arguably more honest while the row is BLOCKED with no runnable vehicle: refuse a config carrying rms_norm_eps by name, which is this tree's own idiom at commandr_registry.cpp:131-138. ~15 lines. A refusal is strictly better than silently wrong numerics.

Not established

Whether any published Cohere checkpoint actually ships rms_norm_eps. No checkpoint was fetched, and the row is BLOCKED on vehicles, so the gap may be latent rather than live.

Read src/vllm/model_executor/models/commandr.cpp and commandr_registry.cpp, starting at LayerNormInto and the layer_norm_eps parsing and refusal logic. Determine whether the intended scope is RMSNorm support or refusing configs containing rms_norm_eps; done means such configs no longer silently use LayerNorm, with coverage for the chosen behavior.