FUSION-DENSE-MIGRATE: the five dgx SACRED paged-engine gates are OWED after the merged-GEMM fold

[localai-bot]

FUSION-DENSE-MIGRATE (#299, PR #324) folded the gate/up of commandr, glm4,
minicpm, minicpm3 and phi3 onto layers::UnquantizedMlpGateUpMethod. The
change is byte-exact by construction and was gated on CPU, but the five SACRED
token-exact gates it touches are checkpoint-gated and dgx-only — on a CPU box
they emit a loud SKIP at 0 assertions, so the empirical confirmation is OWED,
not passed. #299 closes with the fold; this issue is the handle for the runs, so
"PENDING" is tracked somewhere other than prose.

Owed on dgx (GB10), against the pinned oracle

Gate binary	Checkpoint
test_commandr_paged_engine	CohereForCausalLM (Command-R)
test_glm4_paged_engine	Glm4ForCausalLM (GLM-4-9B-0414)
test_minicpm_paged_engine	MiniCPMForCausalLM (MiniCPM-2B-sft-bf16)
test_minicpm3_paged_engine	MiniCPM3ForCausalLM (MiniCPM3-4B)
test_phi3_paged_engine	Phi3ForCausalLM (Phi-4-mini-instruct)

Done when

All five run on dgx at the merged SHA and report their declared verdict (STRICT
where the gate is token-exact, the ratified near-tie band where it is not), with
the result recorded on the FUSION-DENSE-MIGRATE parity-ledger line. A failure
is a real forward divergence and reopens the fold, not this issue.

What already stands without the GPU

• Op-sequence identity: each replaced body was the seam's Apply with M
  spelled T, and every call site passes a DBuf{T,H}.
• test_linear_method — byte-exact seam vs the standalone sequence at M=1 and
  M=4.
• test_dense_gate_up_seam_forward — the four dense folded TUs driven through
  their REAL forward on CPU over synthetic in-memory weights (no checkpoint),
  pinning the gate/up split and half order analytically. minicpm3 is the one
  folded TU that harness does not drive (MLA attention + load-time kv_b_proj
  absorption), so its CPU coverage is the seam unit test only.