npu: check_q4nx_dir calls every gated (attn_output_gate) pack a padded grid — the real d3 pack and all seven 27B qwen3_5 dirs fail --strict #1285

What

tools/check_q4nx_dir.py derives every weight's expected tile grid from config.json and warns when a
container's tile count disagrees. For self_attn.q_proj.weight it expects num_attention_heads × head_dim
rows — but with attn_output_gate the stored q_proj emits q and the output gate, so it holds
2 × that many rows. Every gated pack on this box is therefore reported as a converter-padded grid:

WARN model.layers.3.self_attn.q_proj.weight: holds 512 tiles, the model's dimensions imply 256 -
columns padded 1024 -> 2048 (the 512-column grid); this engine derives the chunk size from the unpadded
grid, so the tensor fails to decode (engine#1097)

and --strict exits 1, i.e. the checker tells an operator that a correct container is broken — the exact
inverse of what it exists for. Found while closing out #796 (the real vllm-sr/d3-lite pack).

Measured (strixhalo, header reads only, no device)

tools/check_q4nx_dir.py <dir> --strict over all 18 Q4NX model directories on this box, before → after:

So the false positive covers 8 of 18 containers here, including the production 27B serve directory, and
nothing that was clean before changes.

The container is right, and the checker is wrong (two independent checks)

• The tensor agrees with 2*QD. d3-lite-conv's model.layers.3.self_attn.q_proj.weight is
  I8 [512, 5120], i.e. 512 tiles of 32×256 = 4096 rows (4096/32 × 1024/256 = 512), against
  QD = 8 × 256 = 2048. The container drops 2 × 2048 = 4096 rows out of q4nx_derive_designs' own
  number: its plan.json layer 3 fused QKV is K=1024 N=5120, and N - 2*KVD = 5120 - 2*512 = 4096 is
  exactly the q|gate pair.
• It decodes and answers. The same container's route returns 10/10 decisions on the array and 10/10
  under NPU_INFER_HOST_GEMM=1 (npu: a real d3 pack (vllm-sr/d3-lite, 0.8B-GDN) answering through /v1/systemone — converter transforms, name map, and the 7-design ELF set #796's rungs 1 and 2), so those six tensors cannot be "failing to decode".

Fix

dims_from_config carries QG = 2*QD when config.json sets attn_output_gate (JSON true or 1 —
the forward reads it with q4nx_config_int), and expected_rc uses QG for q_proj. Two files, 26+/4−:
the tool, a self-test case, and one sentence in docs/npu.md where the checker is described.

Verification

• tools/check_q4nx_dir.py --self-test (ctest q4nx_dir_check) passes, including the new case: a d3
  q_proj of [512, 5120] tiles is quiet under gated dims and still warns under ungated dims, so the case
  can only pass when the gate is what suppresses the warning.
• Tested the test: with qg = qd restored, the self-test fails on
  attn_output_gate did not double q_proj's rows: {'QD': 2048, 'QG': 2048, …}; restored, it passes.
• The sweep above (18 dirs, old vs new script side by side) and --strict rc on the two gated families.
• A freshly converted container (safetensors_to_q4nx.py pack --arch qwen3_5 on the 1.7 GB pack):
  weight bytes byte-identical to the shipped container (sha256 313ac4f9…), verify 320/320 ok, old
  checker 6 WARN / rc 1 → new checker 0 WARN / rc 0.

PR to follow.