Summary

Implements Option<LastTokenStats> field on LayerActivation per SPEC-SHIP-TWO-001 §37.5 Option B + FALSIFY-APR-GGUF-PARITY-007 (apr-vs-gguf-forward-parity-v1.yaml v1.1.0, PR #1107).

What changes

• New LastTokenStats struct mirroring 10 ActivationStats slots, computed only over the last token's slice (hidden_dim or intermediate_dim elements per slot).
• LayerActivation.last_token: Option<LastTokenStats> field, default None for backwards-compat.
• AprTransformer::forward_traced populates last_token via &hidden[(seq_len - 1) * dim..] slicing for all 10 stat slots (attn_norm / qkv / attn_out / ffn_norm / ffn_gate / ffn_up / ffn_silu_gate / ffn_swiglu_inner / ffn_out / output).
• OwnedQuantizedModel::forward_traced populates last_token by cloning existing single-token stats (GGUF already traces only the last token).
• 2 new unit tests pin schema invariants (default-None backwards-compat + populated-count parity).
• 6/6 unit tests PASS.

Live verification (RTX 4090, canonical 7B teacher, prior iteration)

✓ FALSIFY-APR-GGUF-PARITY-007 PASS — count parity restored
APR last_token populated: 28/28
GGUF last_token populated: 28/28

Apples-to-apples ffn_swigl ratios:
  Layer 3 ratio: 1.2154× (Pass)  ← NOT 18.23×!
  ALL 28 layers Pass v1.0.0 ratio gate.

The §27 binding criterion (layer-3 18.23× ratio) was ALMOST ENTIRELY a sample-size artifact — see §38 in PR #1108 for full analysis.

What this does NOT solve

SHIP-007 is still REAL. `apr run` produces "ampiezza = 0.5" gibberish vs GGUF's "2+2 is 4.". But the bug is NOT in layer-3 ffn_swigl. With the misleading binding criterion deprioritized, next-iteration agenda is to bisect SHIP-007 in the autoregressive generation path / KV cache / sampling.

Five-whys (codified in §38.6)

1. Why isn't MODEL-1 inference correct? `apr run` gibberish.
2. Why hasn't §17/§23/§27 chain produced a fix? 18× signal misleading.
3. Why was it artifact? APR all-7-tokens vs GGUF last-token-only.
4. Why didn't earlier reviews catch this? PRs feat(p3-prb): SHIP-007 GGUF forward_traced sub-FFN populate — 4 sub-FFN ActivationStats slots filled #1082+feat(p3-prc): wire apr trace --payload <gguf> to call forward_traced — emits per-layer LayerActivation telemetry #1083 matched API structurally but not semantically.
5. What's the fix? Make both reporters use same sample (this PR).

Plain progress on shipping models

• MODEL-1: this PR makes the FALSIFY-APR-GGUF-PARITY-007 gate satisfiable. With it satisfied, the v1.0.0 ratio gates -001/-002/-003 become CREDIBLE (or, as §38 shows, trivially Pass once measured correctly). Next step: bisect SHIP-007 in autoregressive path.
• MODEL-2: unchanged at val_loss=9.38. Awaiting distill-train impl per docs(p3): apr-cli-distill-train-v1 — contract for missing apr distill train per §35.3 #1097.

Methodology adherence

• Live verification on canonical 7B teacher reproduced stated finding ✓
• Five-whys in §38.6 ✓
• Provable contract referenced (FALSIFY-APR-GGUF-PARITY-007 from v1.1.0) ✓
• Authored in isolated worktree to avoid git-environment race condition that prevented commit in prior iteration ✓

Test plan

• 6/6 unit tests PASS in worktree (`cargo test -p aprender-serve --lib test_layer_activation`)
• PMAT pre-commit gates pass
• Stacks cleanly on main (no dependency on PR docs(ship-007): §37 — APR vs GGUF forward_traced TRACE-CAPTURE-POINT MISMATCH #1105/contract(apr-vs-gguf-forward-parity-v1): v1.0.0 → v1.1.0 — §37 sample-size-parity gate #1107/docs(ship-007): §38 — layer-3 18.23× ratio FALSIFIED as sample-size artifact #1108)

🤖 Generated with Claude Code