feat(M-FFN-GGUF-7-EXT): 28-layer real-teacher chain characterization — saturation confirmed at 1.81×

[noahgift]

Summary

Extends the M-FFN-GGUF-7 5-layer chain test (PR #1548) to ALL 28 layers of the canonical 7B Qwen2.5-Coder-Instruct-Q4_K_M teacher and characterizes the full cumulative-layer pattern — confirming that aggregate drift saturates at ~1.81× even at full model depth despite a single anomalous layer (L24, 442%) that recovers downstream.

• New integration test ffn_gguf_real_teacher_28_layer_chain.rs chains all 28 ffn_down_weight Q4K first super-blocks (Path A vs Path B matvecs)
• Contract amendment v1.12.0 → v1.13.0 records the empirical 28-layer characterization + subsumes the unmade contract bump from PR feat(M-FFN-GGUF-7): multi-layer real-teacher chain — SATURATES at 1.81× (not exponential) #1548
• Total growth factor 1.8103× over 28 layers tracks the M-FFN-GGUF-7 5-layer 1.8081× reference within ±0.1%; saturation dominates aggregate drift; cumulative-layer is NOT a load-bearing amplifier for §27's 1723% magnitude

Empirical highlights

min:                   0.030%   (L2, saturation drop)
max:                 441.978%   (L24, isolated outlier — 1181× jump)
mean:                 16.388%   (skewed by L24)
total growth:          1.8103×  (L27 / L0; matches 5-layer 1.8081×)
saturation events:    13 of 27 transitions (48% drops vs prev)
typical-magnitude:    27 of 28 layers (rel_diff ≤ 10%)

L0-L4 reproduces M-FFN-GGUF-7 5-layer reference values to ≤ 0.001% per layer (validating fixture & chain semantics).

L24 anomaly: weights at L24 first super-block produce a 442% spike (1181× jump from L23), but L25 recovers to 0.271% (0.001× of L24) — chain does NOT enter exponential growth.

Falsifiers

• FALSIFY-FFN-GGUF-016 (5-layer reproduction): DISCHARGED (reference values reproduce exactly)
• FALSIFY-FFN-GGUF-017 (28-layer aggregate growth = 1.81×): DISCHARGED (LIVE pass 2026-05-07, lambda-vector RTX 4090, 26.96s)

Refined §27 magnitude explanation

The 28-layer characterization confirms cumulative-layer is NOT a load-bearing amplifier when measured by aggregate growth (1.81× over 28 layers ≈ 1.81× over 5 layers). Naive growth-factor exponentiation (1.81^(28/5) ≈ 49×) is wrong; real systems saturate via cancellation events.

Updated decomposition:

§27 ≈ M100 × cumulative_saturation × M99
    = 0.428% × 1.81× × 50× ≈ 38.7% drift

vs §27 measured 1723%, residual ~44× now interpretable as per-tensor real-teacher amplitude variation by layer (L24-style anomalies) + 4096-dim std vs M99's 256-dim measurement difference. Resolves when fix Option-A lands.

Test plan

• cargo test -p aprender-serve --release --test ffn_gguf_real_teacher_28_layer_chain --no-run builds clean
• cargo test -p aprender-serve --release --test ffn_gguf_real_teacher_28_layer_chain -- --include-ignored --nocapture LIVE-passes on noah-Lambda-Vector RTX 4090 (26.96s; canonical 7B teacher at /mnt/nvme-raid0/models/ship-two-001/qwen2.5-coder-7b-instruct-q4k.apr)
• pv validate contracts/trace-ffn-sub-block-gguf-v1.yaml exits 0
• cargo clippy -p aprender-serve --release --tests clean for new test file
• rustfmt --check clean for new test file
• Production hot paths byte-unchanged (additive-purity invariant)
• #[ignore]-gated test skips cleanly when canonical teacher absent

🤖 Generated with Claude Code