….0 → v2.59.0 (#1060)

Records the SHIP-007 GQA-7:1 parity bug investigation thread captured
during the 2026-04-25 session as a new §15 of SPEC-SHIP-TWO-001 and
updates the atomic-next-action banner to reference it.

No discharge promotion (coverage tally unchanged at 33+12). This is
investigation-recording, not rule promotion.

What §15 contains:

15.1 Surface symptoms — the two independent observations:
  - apr bench parity gate: CPU argmax 334 vs GPU argmax 8127,
    cosine=−0.005 (anti-correlated, structural divergence)
  - apr qa --json on GGUF: format_parity reports
    GGUF argmax=17 != SafeTensors argmax=59260
  - 370M MODEL-2 training works on the same RTX 4090, so the bug
    is GQA-7:1-specific, not GPU-host-wide.

15.2 Five Whys — traces the surface symptom from the parity-gate
     failure down to the load-bearing edge case: GQA's
     num_heads ≠ num_kv_heads makes layout-then-reshape order
     non-commutative, while MHA (num_heads = num_kv_heads, where
     370M training lives) is invariant under either order.

15.3 Root-cause hypothesis: a GQA-7:1-specific layout-vs-reshape
     ordering bug on K and/or V projections that causes CPU and
     GPU forward to consume the same physical bytes with different
     effective head-axis interpretations, compounding through 28
     transformer blocks into anti-correlated logits.

15.4 Falsifiable next investigation step: a single-tensor Q × K^T
     element-by-element comparison on
     model.layers.0.self_attn.k_proj.weight from the row-major-
     guaranteed APR (SHIP-003 PR #1059), then iterate through V,
     attention scores, weights, output, and o_proj until the
     divergent stage is named. Per feedback_apr_trace_not_eprintln.md,
     this is the proper TraceStep-extension path, not eprintln!.

15.5 Side-bug noted: apr diff --transpose-aware appears not to apply
     the transpose before cosine computation when shapes are [a,b]
     vs [b,a]. Filed as a separate apr-cli ticket. Does not affect
     SHIP-007 root-cause analysis — SafeTensors↔APR same-shape
     comparison via SHIP-003 #1059 confirmed weight-byte parity at
     cos≥0.9999999.

15.6 Blast radius inventory: the remaining 5 MODEL-1 PARTIALs
     (SHIP-002 / 005 / 006 / 007 / 008) all transitively block on
     this single fix. A single root-cause fix discharges all 5
     simultaneously — highest-leverage MODEL-1 work item remaining.

15.7 Methodological note: entire investigation conducted using only
     apr CLI tooling (apr diff, apr qa, apr bench, apr inspect).
     Zero eprintln! injected into forward.rs / ffn_block.rs / CUDA
     kernels. Honors feedback_apr_trace_not_eprintln.md.

Evidence chain:
1. apr diff --values --limit 339 (post-#1058 mmap fix, 192s on
   15GB safetensors / 8GB APR pair) — SafeTensors↔APR cos≥0.9999999.
2. apr diff --values --limit 3 on GGUF↔APR, SafeTensors↔GGUF —
   revealed shape asymmetry: GGUF [in,out] vs APR/SafeTensors
   [out,in].
3. apr qa --json on both APR and GGUF teachers — revealed cross-
   format argmax divergence.
4. SHIP-007 GPU parity gate's existing telemetry — confirmed
   structural divergence.

Methodological consistency with the 6 PR cascade preceding this
amendment: pure stack tooling, contract-backed numbers, drift-
prevention pattern. This commit is documentation only — no Rust
changes, no contract changes — but pins the investigation thread
durably in the spec where future investigators (and the next
multi-PR TraceStep extension effort) will find it.

Spec v2.58.0 → v2.59.0. Atomic-next-action banner updated to point
at §15 as the load-bearing investigation surface.

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

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

…sifier

The §15.4 falsifier from spec v2.59.0 (PR #1060): a CPU vs GPU
incremental_attention_gpu kernel parity test on the canonical
Qwen2.5-Coder-7B shape (NUM_HEADS=28, NUM_KV_HEADS=4, HEAD_DIM=128,
HIDDEN=3584). The peer test `gqa_attention_parity.rs` covers TinyLlama's
GQA-8:1 (NUM_HEADS=32, head_dim=64) — that test passes on RTX 4090 but
doesn't exercise the 7:1 ratio (non-power-of-2 q_per_kv = 7) the
SHIP-007-blocking 7B teacher specifically uses.

Three tests added to
`crates/aprender-serve/tests/qwen2_gqa_7_1_attention_parity.rs`:

1. `ship_007_qwen2_gqa_7_1_head_mapping_property` — pure arithmetic
   check on the GQA-7:1 head mapping (q_head/q_per_kv) for all 28 q_heads.
   Verifies the kernel formula `(q_head * NUM_KV_HEADS) / NUM_HEADS`
   produces identical mapping for the canonical 28:4 ratio.

2. `ship_007_qwen2_gqa_7_1_cpu_gpu_parity_first_token` (#[ignore]) —
   first-token case (no cache). For attention over a single K/V
   position, softmax([single_score]) = [1.0] so output = current_v
   expanded across 4 KV heads to 28 Q heads. CPU reference uses the
   mirror of `cpu_gqa_attention` from the peer test parameterized at
   the Qwen shape. Tolerance: 1e-4 elementwise across 3584 outputs.

3. `ship_007_qwen2_gqa_7_1_cpu_gpu_parity_second_token` (#[ignore]) —
   second-token case with one populated K/V cache position. Tests the
   full attention mechanism with KV cache state. Tolerance: 1e-3
   elementwise (slightly looser to accommodate cumulative FP rounding
   over the 2-position softmax + weighted-sum).

Result on noah-Lambda-Vector RTX 4090 (CUDA 8.9):

  test ship_007_qwen2_gqa_7_1_cpu_gpu_parity_first_token  ... ok
  test ship_007_qwen2_gqa_7_1_cpu_gpu_parity_second_token ... ok
  test ship_007_qwen2_gqa_7_1_head_mapping_property       ... ok

All three pass. **The GQA-7:1 incremental_attention_gpu kernel is NOT
the SHIP-007 root cause.** CPU and GPU outputs are bit-equivalent
(within FP rounding tolerance) for the canonical Qwen2.5-Coder-7B
shape on synthetic inputs.

Materially narrows §15.4's surviving suspect list:
- ✅ Q/K/V head-mapping arithmetic correct (8:1 + 7:1 both pass)
- ✅ Q × K^T per-head correct
- ✅ Softmax-weighted V aggregation correct
- ✅ Scale factor (1/√head_dim) at head_dim=128 correct
- ✅ KV cache state-management correct
- 🟡 Surviving suspects: Q/K/V projection matmul (BEFORE attention),
  o_proj (AFTER attention), RMSNorm, FFN, LM head, multi-layer KV
  cache layout, residual stream propagation.

This test serves as a durable regression guard against the GQA-7:1
attention kernel proper — any future refactor of incremental
attention that breaks 7:1-specific behavior will flip these tests
red on `cargo test --features cuda --release -- --ignored`.

Spec §15.4 (PR #1060) anticipated this test and named the proper
follow-up: a single-tensor matmul parity test on Q/K/V projection
weights from the row-major-correct APR (sha256 a394dd28...0ddeb28,
verified by SHIP-003 PR #1059).

Verification:
  cargo test -p aprender-serve --test qwen2_gqa_7_1_attention_parity \
    --features cuda --release -- --ignored
  → 3 passed; 0 failed; 0 ignored

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

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

…sifier (#1061)

The §15.4 falsifier from spec v2.59.0 (PR #1060): a CPU vs GPU
incremental_attention_gpu kernel parity test on the canonical
Qwen2.5-Coder-7B shape (NUM_HEADS=28, NUM_KV_HEADS=4, HEAD_DIM=128,
HIDDEN=3584). The peer test `gqa_attention_parity.rs` covers TinyLlama's
GQA-8:1 (NUM_HEADS=32, head_dim=64) — that test passes on RTX 4090 but
doesn't exercise the 7:1 ratio (non-power-of-2 q_per_kv = 7) the
SHIP-007-blocking 7B teacher specifically uses.

Three tests added to
`crates/aprender-serve/tests/qwen2_gqa_7_1_attention_parity.rs`:

1. `ship_007_qwen2_gqa_7_1_head_mapping_property` — pure arithmetic
   check on the GQA-7:1 head mapping (q_head/q_per_kv) for all 28 q_heads.
   Verifies the kernel formula `(q_head * NUM_KV_HEADS) / NUM_HEADS`
   produces identical mapping for the canonical 28:4 ratio.

2. `ship_007_qwen2_gqa_7_1_cpu_gpu_parity_first_token` (#[ignore]) —
   first-token case (no cache). For attention over a single K/V
   position, softmax([single_score]) = [1.0] so output = current_v
   expanded across 4 KV heads to 28 Q heads. CPU reference uses the
   mirror of `cpu_gqa_attention` from the peer test parameterized at
   the Qwen shape. Tolerance: 1e-4 elementwise across 3584 outputs.

3. `ship_007_qwen2_gqa_7_1_cpu_gpu_parity_second_token` (#[ignore]) —
   second-token case with one populated K/V cache position. Tests the
   full attention mechanism with KV cache state. Tolerance: 1e-3
   elementwise (slightly looser to accommodate cumulative FP rounding
   over the 2-position softmax + weighted-sum).

Result on noah-Lambda-Vector RTX 4090 (CUDA 8.9):

  test ship_007_qwen2_gqa_7_1_cpu_gpu_parity_first_token  ... ok
  test ship_007_qwen2_gqa_7_1_cpu_gpu_parity_second_token ... ok
  test ship_007_qwen2_gqa_7_1_head_mapping_property       ... ok

All three pass. **The GQA-7:1 incremental_attention_gpu kernel is NOT
the SHIP-007 root cause.** CPU and GPU outputs are bit-equivalent
(within FP rounding tolerance) for the canonical Qwen2.5-Coder-7B
shape on synthetic inputs.

Materially narrows §15.4's surviving suspect list:
- ✅ Q/K/V head-mapping arithmetic correct (8:1 + 7:1 both pass)
- ✅ Q × K^T per-head correct
- ✅ Softmax-weighted V aggregation correct
- ✅ Scale factor (1/√head_dim) at head_dim=128 correct
- ✅ KV cache state-management correct
- 🟡 Surviving suspects: Q/K/V projection matmul (BEFORE attention),
  o_proj (AFTER attention), RMSNorm, FFN, LM head, multi-layer KV
  cache layout, residual stream propagation.

This test serves as a durable regression guard against the GQA-7:1
attention kernel proper — any future refactor of incremental
attention that breaks 7:1-specific behavior will flip these tests
red on `cargo test --features cuda --release -- --ignored`.

Spec §15.4 (PR #1060) anticipated this test and named the proper
follow-up: a single-tensor matmul parity test on Q/K/V projection
weights from the row-major-correct APR (sha256 a394dd28...0ddeb28,
verified by SHIP-003 PR #1059).

Verification:
  cargo test -p aprender-serve --test qwen2_gqa_7_1_attention_parity \
    --features cuda --release -- --ignored
  → 3 passed; 0 failed; 0 ignored

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

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

….0 → v2.62.0

Executed §16.4's first iteration ("apr trace --payload --layer 0 on
both APR and GGUF teachers, bisect through 28 layers") against the
APR teacher's existing per-layer telemetry. The full 28-layer
ffn_out std progression on paiml/qwen2.5-coder-7b-apache-q4k-v1
(prompt "What is 2+2?") shows a 31× discontinuity at layer 3:

  Layer 2: ffn_out std=0.22
  Layer 3: ffn_out std=11.46  ← 31× spike
  Layer 4: ffn_out std=3.84   ← damps in 1 layer (one-off perturbation)
  Median:  ffn_out std=0.5–2.0

The residual stream's output std jumps 0.72 → 11.78 at layer 3 and
stays elevated. Three signals point at layer 3 ffn_out specifically:
(a) magnitude 31× isn't architecture-driven (SHIP-003 PR #1059's
339-tensor cosine sweep proved underlying weights are byte-equivalent
to SafeTensors); (b) damps in one layer (one-off perturbation
pattern, not stable feature); (c) mean shift -0.082 is 100× median
magnitude, suggesting sign-bias defect not magnitude defect.

§17.3 narrows §16.3's four candidates: layer-composition glue in
forward_single_with_scratch at layer 3 FFN is "most likely". Three
new §17.3 candidates added: Q4K dequant under load on 18944-dim FFN;
SiLU numerical stability under SwiGLU `gate * silu(up)`; fused
gate+up matvec dispatch defect (per CLAUDE.md FFN section).

§17.4 specifies sub-layer bisection: emit gate_proj_out, silu(up_proj_out),
gate_proj_out * silu(up_proj_out), down_proj_out separately. Whichever
sub-tensor first shows the 31× std discontinuity vs GGUF path is the
bug site. This requires the §15.5 TraceStep enum extension — now
load-bearing for the fix.

Spec v2.61.0 → v2.62.0. No coverage tally change.

Methodologically: zero eprintln!, zero bash workarounds, third re-use
of `apr trace --payload` primitive without modification (after §15
and §16). Per feedback_apr_trace_not_eprintln.md.

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

….0 → v2.62.0

Executed §16.4's first iteration ("apr trace --payload --layer 0 on
both APR and GGUF teachers, bisect through 28 layers") against the
APR teacher's existing per-layer telemetry. The full 28-layer
ffn_out std progression on paiml/qwen2.5-coder-7b-apache-q4k-v1
(prompt "What is 2+2?") shows a 31× discontinuity at layer 3:

  Layer 2: ffn_out std=0.22
  Layer 3: ffn_out std=11.46  ← 31× spike
  Layer 4: ffn_out std=3.84   ← damps in 1 layer (one-off perturbation)
  Median:  ffn_out std=0.5–2.0

The residual stream's output std jumps 0.72 → 11.78 at layer 3 and
stays elevated. Three signals point at layer 3 ffn_out specifically:
(a) magnitude 31× isn't architecture-driven (SHIP-003 PR #1059's
339-tensor cosine sweep proved underlying weights are byte-equivalent
to SafeTensors); (b) damps in one layer (one-off perturbation
pattern, not stable feature); (c) mean shift -0.082 is 100× median
magnitude, suggesting sign-bias defect not magnitude defect.

§17.3 narrows §16.3's four candidates: layer-composition glue in
forward_single_with_scratch at layer 3 FFN is "most likely". Three
new §17.3 candidates added: Q4K dequant under load on 18944-dim FFN;
SiLU numerical stability under SwiGLU `gate * silu(up)`; fused
gate+up matvec dispatch defect (per CLAUDE.md FFN section).

§17.4 specifies sub-layer bisection: emit gate_proj_out, silu(up_proj_out),
gate_proj_out * silu(up_proj_out), down_proj_out separately. Whichever
sub-tensor first shows the 31× std discontinuity vs GGUF path is the
bug site. This requires the §15.5 TraceStep enum extension — now
load-bearing for the fix.

Spec v2.61.0 → v2.62.0. No coverage tally change.

Methodologically: zero eprintln!, zero bash workarounds, third re-use
of `apr trace --payload` primitive without modification (after §15
and §16). Per feedback_apr_trace_not_eprintln.md.

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

….0 → v2.62.0

Executed §16.4's first iteration ("apr trace --payload --layer 0 on
both APR and GGUF teachers, bisect through 28 layers") against the
APR teacher's existing per-layer telemetry. The full 28-layer
ffn_out std progression on paiml/qwen2.5-coder-7b-apache-q4k-v1
(prompt "What is 2+2?") shows a 31× discontinuity at layer 3:

  Layer 2: ffn_out std=0.22
  Layer 3: ffn_out std=11.46  ← 31× spike
  Layer 4: ffn_out std=3.84   ← damps in 1 layer (one-off perturbation)
  Median:  ffn_out std=0.5–2.0

The residual stream's output std jumps 0.72 → 11.78 at layer 3 and
stays elevated. Three signals point at layer 3 ffn_out specifically:
(a) magnitude 31× isn't architecture-driven (SHIP-003 PR #1059's
339-tensor cosine sweep proved underlying weights are byte-equivalent
to SafeTensors); (b) damps in one layer (one-off perturbation
pattern, not stable feature); (c) mean shift -0.082 is 100× median
magnitude, suggesting sign-bias defect not magnitude defect.

§17.3 narrows §16.3's four candidates: layer-composition glue in
forward_single_with_scratch at layer 3 FFN is "most likely". Three
new §17.3 candidates added: Q4K dequant under load on 18944-dim FFN;
SiLU numerical stability under SwiGLU `gate * silu(up)`; fused
gate+up matvec dispatch defect (per CLAUDE.md FFN section).

§17.4 specifies sub-layer bisection: emit gate_proj_out, silu(up_proj_out),
gate_proj_out * silu(up_proj_out), down_proj_out separately. Whichever
sub-tensor first shows the 31× std discontinuity vs GGUF path is the
bug site. This requires the §15.5 TraceStep enum extension — now
load-bearing for the fix.

Spec v2.61.0 → v2.62.0. No coverage tally change.

Methodologically: zero eprintln!, zero bash workarounds, third re-use
of `apr trace --payload` primitive without modification (after §15
and §16). Per feedback_apr_trace_not_eprintln.md.

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

…ent layer named (#1064)

* docs(ship-007): §17 — layer-3 ffn_out anomaly identified — spec v2.61.0 → v2.62.0

Executed §16.4's first iteration ("apr trace --payload --layer 0 on
both APR and GGUF teachers, bisect through 28 layers") against the
APR teacher's existing per-layer telemetry. The full 28-layer
ffn_out std progression on paiml/qwen2.5-coder-7b-apache-q4k-v1
(prompt "What is 2+2?") shows a 31× discontinuity at layer 3:

  Layer 2: ffn_out std=0.22
  Layer 3: ffn_out std=11.46  ← 31× spike
  Layer 4: ffn_out std=3.84   ← damps in 1 layer (one-off perturbation)
  Median:  ffn_out std=0.5–2.0

The residual stream's output std jumps 0.72 → 11.78 at layer 3 and
stays elevated. Three signals point at layer 3 ffn_out specifically:
(a) magnitude 31× isn't architecture-driven (SHIP-003 PR #1059's
339-tensor cosine sweep proved underlying weights are byte-equivalent
to SafeTensors); (b) damps in one layer (one-off perturbation
pattern, not stable feature); (c) mean shift -0.082 is 100× median
magnitude, suggesting sign-bias defect not magnitude defect.

§17.3 narrows §16.3's four candidates: layer-composition glue in
forward_single_with_scratch at layer 3 FFN is "most likely". Three
new §17.3 candidates added: Q4K dequant under load on 18944-dim FFN;
SiLU numerical stability under SwiGLU `gate * silu(up)`; fused
gate+up matvec dispatch defect (per CLAUDE.md FFN section).

§17.4 specifies sub-layer bisection: emit gate_proj_out, silu(up_proj_out),
gate_proj_out * silu(up_proj_out), down_proj_out separately. Whichever
sub-tensor first shows the 31× std discontinuity vs GGUF path is the
bug site. This requires the §15.5 TraceStep enum extension — now
load-bearing for the fix.

Spec v2.61.0 → v2.62.0. No coverage tally change.

Methodologically: zero eprintln!, zero bash workarounds, third re-use
of `apr trace --payload` primitive without modification (after §15
and §16). Per feedback_apr_trace_not_eprintln.md.

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

* evidence(ship-007): layer-3 ffn_out spike — raw apr trace evidence files

Captures the §17 falsifier evidence as raw artifacts:

  evidence/ship-007-layer-3-anomaly/
  ├── apr-trace-payload-7b-2026-04-26.txt    # 274 lines, all 28 layers
  ├── gguf-trace-payload-7b-2026-04-26.txt   # 34 lines, final decode only
  └── discharge-evidence-v1.json             # JSON summary

Precise measurement: layer-3 ffn_out std = 11.459 / layer-2 ffn_out
std = 0.216 → 53× spike (§17 stated 31×; actual ratio is even more
extreme).

The output residual stream's std jumps 0.7159 (layer 2) → 11.7756
(layer 3) → 25+ (layers 9-19) and never recovers below 13. This
matches the realizar/aprender-serve CLAUDE.md FFN verification
checklist note: "Verify FFN output doesn't cause catastrophic
cancellation" — the layer-3 spike IS that catastrophic
cancellation pattern.

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

---------

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

…→ pass — spec §20 + #1059 evidence — v1.4.0 → v1.5.0

GATE-GPUTRAIN-004 (370M step-time budget < 500ms on RTX 4090) was
marked `verdict: pending` despite its paired falsification test
FALSIFY-GPUTRAIN-005 being DISCHARGED with median 101.30 ms
(20.3% of budget) since 2026-04-24.

This contract bump flips the gate to `verdict: pass` with a
`verdict_basis` field citing both:

1. **FALSIFY-GPUTRAIN-005 evidence** (canonical config seq_len=2048
   batch=1): median 101.30 ms across 25 steps on
   noah-Lambda-Vector RTX 4090 — `evidence/task-132/`.
2. **§20 evidence** (PR #1070, different config seq_len=512):
   median 264.74 ms across 100 steps — `evidence/task-132-residual-b/`.

Both well under the 500ms ceiling. Two evidence files at different
config bands demonstrate budget compliance is robust at this margin.

Contract version v1.4.0 → v1.5.0 (additive metadata, no rule
change). `pv validate`: 0 errors, 0 warnings.

This is a contract-cosmetic flip — GATE-GPUTRAIN-004's underlying
invariant has been satisfied since 2026-04-24; the `verdict: pending`
field was only the gate's own pointer was missing.

References:
- spec §20 (PR #1070): live evidence capture 2026-04-26
- spec §19.4 Residual B: this is the contractual durable verdict
- evidence/task-132/rtx4090-370m-step-budget-and-repro.json
- evidence/task-132-residual-b/cuda-50step-2026-04-26.json

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

…→ pass — spec §20 + #1059 evidence — v1.4.0 → v1.5.0 (#1071)

GATE-GPUTRAIN-004 (370M step-time budget < 500ms on RTX 4090) was
marked `verdict: pending` despite its paired falsification test
FALSIFY-GPUTRAIN-005 being DISCHARGED with median 101.30 ms
(20.3% of budget) since 2026-04-24.

This contract bump flips the gate to `verdict: pass` with a
`verdict_basis` field citing both:

1. **FALSIFY-GPUTRAIN-005 evidence** (canonical config seq_len=2048
   batch=1): median 101.30 ms across 25 steps on
   noah-Lambda-Vector RTX 4090 — `evidence/task-132/`.
2. **§20 evidence** (PR #1070, different config seq_len=512):
   median 264.74 ms across 100 steps — `evidence/task-132-residual-b/`.

Both well under the 500ms ceiling. Two evidence files at different
config bands demonstrate budget compliance is robust at this margin.

Contract version v1.4.0 → v1.5.0 (additive metadata, no rule
change). `pv validate`: 0 errors, 0 warnings.

This is a contract-cosmetic flip — GATE-GPUTRAIN-004's underlying
invariant has been satisfied since 2026-04-24; the `verdict: pending`
field was only the gate's own pointer was missing.

References:
- spec §20 (PR #1070): live evidence capture 2026-04-26
- spec §19.4 Residual B: this is the contractual durable verdict
- evidence/task-132/rtx4090-370m-step-budget-and-repro.json
- evidence/task-132-residual-b/cuda-50step-2026-04-26.json

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

…efined hypothesis ALSO FALSIFIED

Authors a third lib-only falsifier (FALSIFY-FFN-GGUF-006) in
apr_transformer::helpers::determinism_tests:

  falsify_ffn_gguf_006_simd_vs_scalar_reduction_order_byte_identity

Test runs APR's simd_dot_f32_avx2 (AVX2 8-wide FMA) and APR's
scalar fallback (iter().zip().map(*).sum()) on the same canonical
synthetic input, compares bit patterns via f32::to_bits().

EMPIRICAL RESULT (2026-05-06): both paths produce BYTE-IDENTICAL
output 0x44191e70 = 612.4756. Asserted as regression-test
invariant.

This FALSIFIES the refined H2a' hypothesis at the SIMD-vs-scalar
level. The cumulative APR↔GGUF drift cannot be explained by APR's
SIMD vs APR's scalar path differing on this class of f32 inputs.

SECOND HYPOTHESIS FALSIFICATION IN ONE SESSION:
- §28 (parallel-reduction non-determinism, M91): FALSIFIED
- H2a' (SIMD-vs-scalar reduction-order, this PR): FALSIFIED

NEW REFINED HYPOTHESIS H2d (post-second-falsification):
The bit-level difference between APR and GGUF must come from one
of:

H2d.1: Per-block dequant boundaries differ between APR's whole-row
       F32 reduction and GGUF's Q4K-super-block-wise reduction
H2d.2: APR's F32 weights differ at bit level from a true
       dequantization of the GGUF Q4K bytes (despite SHIP-003 PR
       #1059 cos≥0.9999999 weight invariance)
H2d.3: GGUF's intermediate Q8K activation quantization rounds
       activations to ~7-bit precision differently than APR's
       full-F32 path

Each H2d.x is a separate falsifier candidate.

Next M-FFN-GGUF-4 step (c) deliverable: H2d.2 is most directly
testable autonomously — load APR F32 weights + GGUF Q4K bytes for
same tensor, dequantize Q4K via APR's dequant routine, compare
element-wise. If bit-level differ, H2d.2 confirmed.

Contract amendment: trace-ffn-sub-block-gguf-v1 v1.2.0 → v1.3.0.

Status promotions:
- FALSIFY-FFN-GGUF-006: NEW → DISCHARGED (test passes after flip)
- M-FFN-GGUF-4 step (b): PENDING → SHIPPED

Step (c) remains PENDING — narrowed scope to H2d.{1,2,3}.

Production hot paths byte-unchanged.

`pv validate` 0/0; 3 lib tests pass.

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

…efined hypothesis ALSO FALSIFIED (#1536)

Authors a third lib-only falsifier (FALSIFY-FFN-GGUF-006) in
apr_transformer::helpers::determinism_tests:

  falsify_ffn_gguf_006_simd_vs_scalar_reduction_order_byte_identity

Test runs APR's simd_dot_f32_avx2 (AVX2 8-wide FMA) and APR's
scalar fallback (iter().zip().map(*).sum()) on the same canonical
synthetic input, compares bit patterns via f32::to_bits().

EMPIRICAL RESULT (2026-05-06): both paths produce BYTE-IDENTICAL
output 0x44191e70 = 612.4756. Asserted as regression-test
invariant.

This FALSIFIES the refined H2a' hypothesis at the SIMD-vs-scalar
level. The cumulative APR↔GGUF drift cannot be explained by APR's
SIMD vs APR's scalar path differing on this class of f32 inputs.

SECOND HYPOTHESIS FALSIFICATION IN ONE SESSION:
- §28 (parallel-reduction non-determinism, M91): FALSIFIED
- H2a' (SIMD-vs-scalar reduction-order, this PR): FALSIFIED

NEW REFINED HYPOTHESIS H2d (post-second-falsification):
The bit-level difference between APR and GGUF must come from one
of:

H2d.1: Per-block dequant boundaries differ between APR's whole-row
       F32 reduction and GGUF's Q4K-super-block-wise reduction
H2d.2: APR's F32 weights differ at bit level from a true
       dequantization of the GGUF Q4K bytes (despite SHIP-003 PR
       #1059 cos≥0.9999999 weight invariance)
H2d.3: GGUF's intermediate Q8K activation quantization rounds
       activations to ~7-bit precision differently than APR's
       full-F32 path

Each H2d.x is a separate falsifier candidate.

Next M-FFN-GGUF-4 step (c) deliverable: H2d.2 is most directly
testable autonomously — load APR F32 weights + GGUF Q4K bytes for
same tensor, dequantize Q4K via APR's dequant routine, compare
element-wise. If bit-level differ, H2d.2 confirmed.

Contract amendment: trace-ffn-sub-block-gguf-v1 v1.2.0 → v1.3.0.

Status promotions:
- FALSIFY-FFN-GGUF-006: NEW → DISCHARGED (test passes after flip)
- M-FFN-GGUF-4 step (b): PENDING → SHIPPED

Step (c) remains PENDING — narrowed scope to H2d.{1,2,3}.

Production hot paths byte-unchanged.

`pv validate` 0/0; 3 lib tests pass.

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