Context

After the M-GPU-MOE-1.x cascade closure (M51-M85, see paiml/claude-code-parity-apr companion-repo spec § Sub-extension 2 status as of M85-M87), the wgpu sibling path remains incomplete.

Stub merged at M54 (#1485 squash 5a27bb892, 3-commit bundle including OwnedQuantizedModelWgpu at crates/aprender-serve/src/gguf/wgpu_backend/mod.rs).

Open work

• M-GPU-MOE-2.1: expert_swiglu_wgpu + moe_ffn_forward_layer_wgpu helpers — blocked on trueno-gpu wgpu surface authoring (QuantizeKernel + GemmKernel compute pipelines).
• M-GPU-MOE-2.2: full forward integration mirroring forward_qwen3_moe_cuda (M51 feat(aprender-serve): forward_qwen3_moe_cuda full integration — M-GPU-MOE-1.1.2 #1477 squash dc6f94d3b).
• M-GPU-MOE-2.3: heavy --include-ignored cosine ≥ 0.99 vs CPU LAZY-FUSED-MATVEC parity test (FALSIFY-QW3-MOE-GPU-PARITY-001 wgpu sibling).

Test scaffold already authored

• crates/aprender-serve/tests/qwen3_moe_wgpu_parity.rs (M54 test(aprender-serve): qwen3_moe_wgpu_parity — M-GPU-MOE-2.3 cosine ≥0.99 falsifier (wgpu) #1488 squash 10cc7ad41)

Acceptance

Heavy --include-ignored runs PASS on:

• Apple Silicon Metal
• AMD Vulkan
• Intel ARC

Cross-refs

• Contract: aprender-contracts/contracts/qwen3-moe-forward-gpu-v1.yaml v1.7.0 ACTIVE_ALGORITHM_LEVEL (M86 contract(qwen3-moe-forward-gpu-v1): v1.6.0 → v1.7.0 — DRAFT → ACTIVE_ALGORITHM_LEVEL #1530 squash 65bc42577)
• Companion-repo: paiml/claude-code-parity-apr § Sub-extension 2 deliverable 3 (PENDING)
• Related: GH-560: Wire wgpu fallback into batch inference path #560 (wgpu fallback into batch inference path), apr serve run: Q4K APR GPU hangs on large MoE models (30B) #471 (Q4K APR GPU hangs on large MoE models)