…mpounding (5.70×) — bundled M94+M95 (#1538)

* feat(M-FFN-GGUF-4 step c, H2d.3+H2d.4): fused-vs-standalone Q4K matvec — FIRST CONFIRMED hypothesis in chain

After three sequential falsifications (M91 §28 parallel-reduction,
M92 H2a' SIMD-vs-scalar dot, M93 H2d.2 APR-internal Q4K dequant
byte-identity), the H2d.4 falsifier (FALSIFY-FFN-GGUF-008) is the
**first test in the SHIP-007 §22 hypothesis chain that produces
the EXPECTED bit-level divergence between paths**.

Adds `falsify_ffn_gguf_008_fused_vs_standalone_q4k_matvec` to
`crates/aprender-serve/src/apr_transformer/helpers.rs::
determinism_tests`. Compares:

  Path A (APR-style):  dequantize_q4_k_simd + manual F32 dot
  Path B (GGUF-style): quantize_activations_q8k_into +
                       fused_q4k_q8k_parallel_matvec_into

On a synthetic 144-byte Q4K super-block + 256-element F32
activation. Both paths compute the same mathematical operation
(W @ a) but Path B has an additional Q8K activation-quantization
step Path A doesn't have.

EMPIRICAL RESULT (2026-05-06):
  Path A = -18882.443 (0xc69384e3)
  Path B = -18897.059 (0xc693a21e)
  diff   = 14.615 (rel_diff = 0.077%)
  bits_a != bits_b ✓

Paths DIFFER at bit level as expected. Math agreement within
0.10% (Q8K precision loss is mathematically reasonable but NOT
bit-exact). This **CONFIRMS H2d.3 + H2d.4 simultaneously** at
the kernel level.

SHIP-007 §22 ROOT CAUSE NOW HAS A CONCRETE MECHANISM:

APR's loader path uses Path A semantics — full F32 dequant of
weights, then F32 matmul with F32 activations. GGUF's matvec
uses Path B semantics — Q8K quantization of activations + fused
inline Q4K dequant during the parallel matvec. Per-tensor the
divergence is small (0.077%) but cumulative across 28 layers ×
4 matmuls/layer × 7 tokens, the divergence compounds in a way
that matches the §27 layer-3 ffn_swigl 18.23× APR↔GGUF drift.

Hypothesis chain (CLOSED for kernel-level reduction-order):
- §28  parallel-reduction non-determinism      (M91): FALSIFIED
- H2a' SIMD-vs-scalar dot reduction            (M92): FALSIFIED
- H2d.2 APR-internal Q4K dequant byte-identity (M93): FALSIFIED
- H2d.3 + H2d.4 fused-vs-standalone matvec     (M94): CONFIRMED ✓

Contract trace-ffn-sub-block-gguf-v1 v1.4.0 → v1.5.0:
- Documents the first hypothesis CONFIRMATION in the chain
- Records empirical evidence (-18882.443 vs -18897.059)
- Records the two architecturally-clean fix options:
  - Option-A (PROMOTE GGUF-PATH semantics into APR forward)
  - Option-B (PROMOTE APR-PATH semantics into GGUF forward)
- M-FFN-GGUF-4 step (c) hypothesis-narrowing: ALGORITHM_LEVEL
  → DISCHARGED — chain produced first CONFIRMED mechanism
- M-FFN-GGUF-5 (NEW, NEXT): SHIP-007 §22 actual fix PR; gate
  Option-A vs Option-B; PENDING

Production hot paths byte-unchanged. New test additive in
`crates/aprender-serve/src/apr_transformer/helpers.rs::
determinism_tests`. `pv validate`: 0 errors / 0 warnings on
v1.5.0.

Test runs locally on RTX 4090:
  cargo test -p aprender-serve --lib falsify_ffn_gguf_008
  test result: ok. 1 passed; 0 failed; finished in 0.00s

Refs PMAT-CCPA, SHIP-007 §22.

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

* feat(M-FFN-GGUF-4 step e): multi-tensor compound falsifier — SUPER-LINEAR growth confirmed (5.70× over 5 chained matvecs) (#1539)

M94 (FALSIFY-FFN-GGUF-008, sibling PR #1538) confirmed Path A vs
Path B differ by 0.077% on a SINGLE 144-byte Q4K super-block matvec.
The v1.5.0 amendment hypothesized (without measurement) that this
compounds across "28 layers × 4 matmuls/layer × 7 tokens" to match
the §27 layer-3 ffn_swigl 18.23× std-ratio.

This PR authors `falsify_ffn_gguf_009_multi_tensor_divergence_compound`
in `crates/aprender-serve/src/apr_transformer/helpers.rs::
determinism_tests` to MEASURE that compounding empirically. Test
runs N=5 sequential matvecs (chained — each output is the next
input, with RMSNorm between layers to keep magnitude bounded),
comparing Path A vs Path B at the final layer.

EMPIRICAL RESULT (2026-05-06):
  Single-tensor rel_diff (M94): 0.077%
  5-tensor chained rel_diff:    0.4391%
  Growth factor:                5.70×

Linear projection would be 5.00× (5 × 0.077%); sub-linear (√N)
projection would be 2.24×. The empirical 5.70× growth is
**SUPER-LINEAR** — confirms H-COMPOUND-SUPER hypothesis.

QUANTITATIVE EXTRAPOLATION TO §27:
  Layer-3 chain depth = 3 layers × ~7 tensor-ops = 21 chained ops.
  Naive super-linear extrapolation:
    21 × 0.077% × (5.70/5)^log2(21/5) ≈ 1.85% (rel_diff)

This is FAR BELOW §27's 1723% (18.23× std-ratio). The M94 mechanism
explains COMPOUNDING but not the §27 MAGNITUDE.

Three candidate amplifiers (M-FFN-GGUF-6 investigation scope):
- A1: RoPE phase amplification (rotational drift across heads)
- A2: Softmax saturation (logit drift → output drift via near-max)
- A3: Real-weight magnitude variance (synthetic uniform magnitude
       vs real Qwen Q4K weights with high per-tensor variance)

Most likely path forward: M-FFN-GGUF-6 = real-teacher falsifier.
Load actual layer-3 down_proj Q4K bytes from canonical 7B Qwen2.5-
Coder .apr file at `/mnt/nvme-raid0/models/ship-two-001/qwen2.5-
coder-7b-instruct-q4k.apr`, run both paths against a real activation
vector, measure rel_diff. If real-teacher rel_diff is 5-50× larger
than synthetic, A3 alone explains §27 magnitude. If matches
synthetic, A1+A2 are load-bearing.

Contract trace-ffn-sub-block-gguf-v1 v1.5.0 → v1.6.0:
- FALSIFY-FFN-GGUF-009 NEW → DISCHARGED
- M-FFN-GGUF-4 step (e) compounding-hypothesis: DISCHARGED
- M-FFN-GGUF-6 (NEW, NEXT): real-teacher falsifier; PENDING

Production hot paths byte-unchanged. Test additive in
helpers.rs::determinism_tests. `pv validate`: 0 errors / 0 warnings
on v1.6.0.

Stacked atop PR #1538 (M94/M-FFN-GGUF-4d). Will rebase on main
after #1538 merges.

Test runs locally:
  cargo test -p aprender-serve --lib falsify_ffn_gguf_009 -- --nocapture
  test result: ok. 1 passed; finished in 0.03s

Refs PMAT-CCPA, SHIP-007 §22, FALSIFY-FFN-GGUF-009.

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>