Adds `crates/aprender-serve/examples/diag_q4k_dequant_cpu_vs_gpu.rs`
that loads the canonical 7B teacher via `OwnedQuantizedModel::from_apr`
(same path apr run uses) and:

1. Classifies QKV storage as Fused vs Separate.
2. Dequantizes Q/K/V via `dequantize_q4_k` (same function GPU upload uses).
3. Reports per-projection stats + first 8 elements + sanity bounds.

Live result on canonical 7B teacher (RTX 4090):

  Layer 0 QKV is SEPARATE
  q tensor: in_dim=3584 out_dim=3584 qtype=12 (Q4K)
  k tensor: in_dim=3584 out_dim=512  qtype=12 (Q4K)
  v tensor: in_dim=3584 out_dim=512  qtype=12 (Q4K)

  Q-projection: mean=-0.000001 std=0.021468 [-0.6056, +0.5965]
  K-projection: mean=-0.000039 std=0.032039 [-0.2383, +0.2478]
  V-projection: mean=0.000023 std=0.010917 [-0.0840, +0.0845]

  All within sane bounds (max |Q|=0.61, |K|=0.25, |V|=0.08).

Findings:
- H1 (Q4K dequant correctness): values look sane; refuted as primary cause
  unless the GPU upload path dequantizes DIFFERENTLY from `dequantize_q4_k`
  (next-iter check).
- H2 (Fused-QKV slicing bug at wgpu_adapter.rs:56-62): NOT exercised on this
  model — QKV is SEPARATE, not Fused. Refuted for canonical 7B teacher.

Remaining hypothesis: H3 (wgpu vs CUDA dispatch interplay or wgpu-internal
defect). Notable from `apr run` log: `[PMAT-333] Dequantized 337 weights,
28282.5 MB F32` — 28 GB total F32 cache; only lm_head (2180 MB) falls back
to CPU; other weights presumably go to wgpu/CUDA. RTX 4090 has 24 GB VRAM,
so wgpu buffer management must spill or evict — possible defect surface.

Per CLAUDE.md MANDATORY GPU TESTING + LAYOUT-001/002 row-major mandate,
the next bisection step is to dump actual GPU-uploaded F32 weights and
compare element-wise vs the values printed by this diag. If values match,
the dequant path is correct and the bug is in matmul/RoPE/attention OR
in the upload/dispatch chain (wgpu buffer mgmt, layout transpose, etc.).

Five-whys (consistent with §40.5):
1. Why is `apr run` wrong on GPU? "ampiezza = 1" instead of correct math.
2. Why? GPU dispatch corrupts forward computation.
3. Why? Per §40.4 it's NOT FP8 and per this diag it's likely NOT dequant-
   value correctness either.
4. Why? Likely the upload/dispatch chain (wgpu buffer mgmt, layout).
5. What's the fix? Next-iter: dump GPU-uploaded weights and diff vs this
   diag's reference values; OR add wgpu-disable env var if exists.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>