…tion cascade STARTED (#1454)

* contract(qwen3-moe-forward-gpu-v1): v1.0.0 DRAFT — scaffold for P0 GPU MoE forward path

Per claude-code-parity-apr POC M49 priority elevation 2026-05-04, the
GPU MoE forward path is now P0 / HIGHEST PRIORITY. Per CLAUDE.md
"NEVER write code before writing a provable contract" — this is the
contract scaffold (M-stage M-GPU-MOE-0 in the contract's
implementation_stages).

Why P0
======

  - CPU LAZY-FUSED-MATVEC produces correct output but at ~30 tok/s on
    Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.
  - Dense GPU Q4_K (Qwen2.5-Coder-7B) on RTX 4090 cuBLAS: 225-440 tok/s.
  - MoE inference is ~10× slower than dense, making the spec-prescribed
    default Qwen3-Coder model production-infeasible at ~30 tok/s.
  - The companion's action-stream parity machinery (CCPA-001..013, all
    DISCHARGED) cannot be exercised at production cadence — every
    `apr code` invocation hits the 30 tok/s wall.

What this contract specifies
============================

  metadata.kind: kernel
  status:        DRAFT
  scope:         crates/aprender-serve/src/gpu/{forward_qwen3_moe_gpu,
                   scheduler/moe_dispatch}.rs +
                 crates/aprender-compute/src/gpu/moe_kernels.rs (TBD)

  equations:
    - moe_forward_one_layer_gpu  (mirrors v1 CPU equation, +cosine-vs-CPU
                                   invariant, +CudaExecutor::new(0).is_ok()
                                   precondition)
    - gpu_throughput_target      (≥150 tok/s on RTX 4090 over 128-tok
                                   median window, ≥5x CPU baseline)

  proof_obligations: 7
    AC_GPU_MOE_001  cosine ≥ 0.99 vs CPU LAZY-FUSED-MATVEC
    AC_GPU_MOE_002  router weights sum to 1.0 ± 1e-6
    AC_GPU_MOE_003  output dimensions preserved
    AC_GPU_MOE_004  output finite (no NaN/Inf)
    AC_GPU_MOE_005  cosine ≥ 0.99 vs HF FP16 (inherits from v1)
    AC_GPU_MOE_006  ≥150 tok/s on RTX 4090
    AC_GPU_MOE_007  VRAM utilization ≤ 95% of 24 GB

  falsification_tests: 7
    FALSIFY-QW3-MOE-GPU-001          baseline (no GPU symbol)
    FALSIFY-QW3-MOE-GPU-PARITY-001   M-GPU-MOE-1 cosine vs CPU
    FALSIFY-QW3-MOE-GPU-PARITY-002   M-GPU-MOE-1 cosine vs HF FP16
    FALSIFY-QW3-MOE-GPU-INVARIANTS-001 router/shape/finite
    FALSIFY-QW3-MOE-GPU-DETERMINISM-001 byte-identical reruns same seed
    FALSIFY-QW3-MOE-GPU-THROUGHPUT-001 ≥150 tok/s
    FALSIFY-QW3-MOE-GPU-MEMORY-001     ≤ 95% VRAM

  kani_harnesses: 2
    KANI-QW3-MOE-GPU-001  router weights sum (AC_GPU_MOE_002)
    KANI-QW3-MOE-GPU-002  output shape preservation (AC_GPU_MOE_003)

  qa_gate: F-QW3-MOE-GPU-001 (5 named checks, falsification = swap
    quantized for EAGER FP32 → guaranteed OOM on 24 GB VRAM)

Implementation stages
=====================

  M-GPU-MOE-0  This contract scaffold                        SHIPPED
  M-GPU-MOE-1  CUDA kernel + cosine-vs-CPU parity gate       PENDING
  M-GPU-MOE-2  wgpu fallback (CLAUDE.md backend-agnostic)    PENDING
  M-GPU-MOE-3  Throughput ≥150 tok/s + VRAM ≤ 95%            PENDING

When all 3 PENDING stages discharge, status flips DRAFT →
ACTIVE_RUNTIME (matches qwen3-moe-forward-v1 v1 convention).

Verification
============

  $ pv validate contracts/qwen3-moe-forward-gpu-v1.yaml
  0 error(s), 0 warning(s)
  Contract is valid.

Refs claude-code-parity-apr POC M49 (P0 elevation, 2026-05-04)
Refs claude-code-parity-apr POC R10 (risk row mirror)
Refs qwen3-moe-forward-v1 v1.4.0 ACTIVE_ALGORITHM_LEVEL (CPU sibling)
Refs apr-cpu-vs-gpu-output-parity-v1 (CPU↔GPU parity discipline)
Refs arXiv:2305.18398 Dao FlashAttention-2
Refs arXiv:2305.05176 Aminabadi DeepSpeed-MoE
Refs arXiv:2101.03961 Fedus Switch Transformers

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

* spec(ship-two-models): v2.91.0 → v2.92.0 — §47 SHIP-007 layer-0 attention bisection cascade STARTED

After §46 declared the v0.32.0 cut HOLD-gated on SHIP-007 layer-0 attention,
the §46.7(a) follow-up cascade kicked off with three PRs in flight (#1450 +
#1451 + #1452).

## What §47 records

| Subsection | Content |
|---|---|
| 47.1 | 8-step cascade roadmap (this amendment captures steps 1-3) |
| 47.2 | What landed in PRs #1450 + #1451 + #1452 |
| 47.3 | **Toyota Way correction in detail** — v1.0.0 → v1.1.0 mid-cascade |
| 47.4 | Pre-existing parent contract drift (QPostRope/KPostRope unwired) |
| 47.5 | Net effects (ship %, coverage tally, pending merges) |
| 47.6 | Open follow-ups (5-step ranked priority list) |
| 47.7 | Five Whys (why amend at 3 PRs, why split §47/§48, etc.) |
| 47.8 | Spec amendment cadence preserved (§41 → §47, 7 amendments) |

## Cascade roadmap

| # | PR | What | Discharge status |
|---|----|------|-------|
| 1 | #1450 | Contract `trace-attn-sub-stages-v1.yaml` v1.1.0 PROPOSED | 5 falsifiers algorithm-bound |
| 2 | #1451 | Enum extension: 2 new SaveTensorStage variants | FALSIFY-ATTN-SUB-001 PARTIAL_ALGORITHM_LEVEL |
| 3 | #1452 | Research evidence note | No falsifier flip |
| 4 | (next) | forward_traced_with_plan wires 4 sub-stages | FALSIFY-ATTN-SUB-002 + drift fix |
| 5 | (next) | apr diff --values recognizes new stages | FALSIFY-ATTN-SUB-003 |
| 6 | (next) | HF FP16 oracle script extension | unblocks FALSIFY-ATTN-SUB-004 |
| 7 | (next) | Live RTX 4090 bisection | FALSIFY-ATTN-SUB-004 → DISCHARGED |
| 8 | (next) | SHIP-007 root-cause fix | unblocks MODEL-1 GPU |

§47 captures the first 3 (scaffold). §48+ will capture later steps.

## Toyota Way correction (mid-cascade)

v1.0.0 of `trace-attn-sub-stages-v1.yaml` was the day's first defect. It
claimed 5 new SaveTensorStage variants needed; live source inspection
(per `feedback_no_guessing.md`) showed 3 already existed. v1.1.0 corrected
to 2 truly-new variants + added the 9-element `bisection_chain_layer_0`
equation. Cost-of-defect paid at the contract layer (cheapest place); no
code rolled back.

## Pre-existing parent contract drift

Researching the wire-plan for FALSIFY-ATTN-SUB-002 surfaced a drift in
`apr-cli-trace-save-tensor-v1.yaml` v1.4.0 (FUNCTIONAL): `QPostRope` +
`KPostRope` are in the enum but have NO `emit()` calls in
`forward_traced_with_plan`. A user passing `--save-tensor q_post_rope`
gets a clean exit with no file written — silent failure.

Per `feedback_toyota_way_all_defects.md`: all defects are mine. The
next-cycle FALSIFY-ATTN-SUB-002 PR will close this drift as a free
side-effect by wiring the 2 missing stages alongside the 2 new ones.

## Net effects

- Spec v2.91.0 → **v2.92.0**
- Coverage tally: unchanged this cycle (5 new PARTIAL_ALGORITHM_LEVEL
  slots will increment when PR #1450 lands the YAML)
- MODEL-1 ship %: unchanged at 91% (cascade is scaffold; ship % moves
  at FALSIFY-ATTN-SUB-004 LIVE DISCHARGE in a future cycle)
- MODEL-2 ship %: unchanged at 57%

## Five Whys (compressed)

1. Why amend at 3 PRs? §41-§46 cadence is "one amendment per ≥3-PR cycle"
2. Why split §47/§48? Toyota Way correction is worth pinning
3. Why pin parent drift here, not amend the parent contract? Drift fix
   lands in next-cycle implementation PR; §47 just records it
4. Why no FALSIFY-ATTN-SUB-002 in this cycle? Single-piece flow;
   stacked PRs slow merge throughput
5. Why no parent-contract bump now? Bump requires wire fix landing first
   (FUNCTIONAL claim) — cleaner to bump in next-cycle PR

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

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

---------

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

…ith_plan — FALSIFY-ATTN-SUB-002 (#1455)

* contract(trace-attn-sub-stages-v1): scaffold layer-0 attention bisection (5 new SaveTensorStage variants)

Authors a new provable-contract `trace-attn-sub-stages-v1.yaml` v1.0.0 PROPOSED
that pre-commits to the schema for extending `SaveTensorStage` with FIVE new
intermediate attention-block sub-stages so SHIP-007 layer-0 attention divergence
can be bisected element-wise against the HF FP16 oracle (PR #1423).

## Why now (per spec §46.7)

Spec v2.91.0 §46.7 ranked SHIP-007 layer-0 attention bisection as the highest-
leverage MODEL-1 follow-up. Memory `2026-05-03 SHIP-007 finding`:

- cos(APR.attn_norm, HF.attn_norm) = 0.99999995  ✓ (correct)
- cos(APR.attn_out,  HF.attn_out)  = 0.9966      ✗ (wrong)

The bug is somewhere INSIDE the attention block. The existing
`SaveTensorStage` enum has only `QkvMatmul` between `AttnNorm` and `AttnOut` —
too coarse to localize.

## What this contract pins

5 new variants, in computation order inside the attention block:

| New stage | What it captures |
|---|---|
| `QPostRope`   | Q after RoPE (post Q-projection + RoPE rotate) |
| `KPostRope`   | K after RoPE (GQA: shared across head groups) |
| `AttnScores`  | Q·Kᵀ / sqrt(head_dim), pre-softmax |
| `AttnSoftmax` | softmax(scores + causal_mask) |
| `AttnVOut`    | softmax · V (pre output O-projection) |

Capture order: `QkvMatmul → QPostRope → KPostRope → AttnScores → AttnSoftmax → AttnVOut → AttnOut`

## Falsifiers (5)

| ID | What it predicts | Status |
|---|---|---|
| FALSIFY-ATTN-SUB-001 | 5 new variants exist; existing 14 preserved byte-identical | PARTIAL_ALGORITHM_LEVEL |
| FALSIFY-ATTN-SUB-002 | `forward_traced_with_plan` threads them in canonical order | PARTIAL_ALGORITHM_LEVEL |
| FALSIFY-ATTN-SUB-003 | `apr diff --values` recognizes APRT files for the 5 stages | PARTIAL_ALGORITHM_LEVEL |
| FALSIFY-ATTN-SUB-004 | Bisection narrows SHIP-007 to ONE specific sub-stage | BLOCKER_FIXTURE_ABSENT |
| FALSIFY-ATTN-SUB-005 | Capture is purely additive (token output byte-identical) | PARTIAL_ALGORITHM_LEVEL |

FALSIFY-ATTN-SUB-004 is the load-bearing one — it is the predicate that must
be falsified to actually pinpoint the SHIP-007 sub-stage. Marked
BLOCKER_FIXTURE_ABSENT because live discharge requires (i) the 5 new stages
implemented, (ii) HF FP16 oracle extended to capture them, (iii) live diff on
RTX 4090. This contract pins the gate; the implementation cascade follows.

## Five Whys

1. **Why a new contract instead of extending `apr-cli-trace-save-tensor-v1`?**
   The parent contract is FUNCTIONAL (v1.4.0); extending it would re-open it.
   Mirrors the `trace-ffn-sub-block-v1` SHIP-007 layer-3 prior art (#1083) —
   sub-block contracts are siblings of the parent, not amendments.

2. **Why pin the schema before implementation?**
   Per `feedback_apr_trace_not_eprintln.md`: "Missing TraceStep granularity →
   extend the enum behind a contract." Contract-first preserves the audit
   chain spec § → contract → implementation PRs → live discharge.

3. **Why these 5 stages and not 3 or 7?**
   The 5 capture points bracket every numerically distinct intermediate
   inside attention: pre-RoPE (QkvMatmul exists), Q post-rope, K post-rope,
   scores (Q·Kᵀ), softmax (post-mask + softmax), V·softmax (pre O-proj).
   Adding sub-stages of these (e.g., separate Q vs K matmul outputs) is
   premature — let the bisection localize first, then refine if needed.

4. **Why mark FALSIFY-ATTN-SUB-004 as BLOCKER_FIXTURE_ABSENT and not PARTIAL?**
   PARTIAL_ALGORITHM_LEVEL means an algorithm reference exists today.
   ATTN-SUB-004's discharge requires LIVE evidence + the HF FP16 oracle
   extension; today neither exists. BLOCKER honestly classifies the gap;
   matches `apr-cli-distill-train-v1` TRAIN-009 precedent (§43, PR #1443).

5. **Why is this not just SHIP-007's fix itself?**
   Fixing SHIP-007 needs to know WHICH sub-stage is wrong. This contract
   delivers the *measurement instrument* that pinpoints the sub-stage; the
   fix is the next PR cascade after that pin lands.

## Net effects

- New contract `trace-attn-sub-stages-v1.yaml` v1.0.0 PROPOSED, 5 falsifiers.
- `pv validate contracts/trace-attn-sub-stages-v1.yaml` exits 0.
- MODEL-1 ship %: unchanged at 91% (this is contract scaffold; no falsifier flips).
- MODEL-2 ship %: unchanged at 57%.
- Coverage tally: unchanged this PR (4 PARTIAL + 1 BLOCKER added but contract
  is new — they count once it''s wired into the §-amendment chain).
- Unblocks the next PR cascade: enum extension + forward_traced threading +
  apr diff recognition + HF FP16 oracle extension → FALSIFY-ATTN-SUB-001..005
  algorithm-bind → live RTX 4090 bisection → ATTN-SUB-004 DISCHARGE.

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

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

* contract(trace-attn-sub-stages-v1): v1.0.0 → v1.1.0 — Toyota Way correction (only 2 new variants needed, not 5)

## What's wrong with v1.0.0

v1.0.0 (commit 475dec3) claimed FIVE new SaveTensorStage variants
were needed for the SHIP-007 layer-0 attention bisection:
QPostRope, KPostRope, AttnScores, AttnSoftmax, AttnVOut.

Empirical inspection of `crates/aprender-serve/src/inference_trace/save_tensor_stage.rs`
shows THREE of those five ALREADY EXIST in the parent contract
`apr-cli-trace-save-tensor-v1.yaml` v1.4.0 FUNCTIONAL:

- `QPostRope`  — already in enum (line 47)
- `KPostRope`  — already in enum (line 49)
- `Attention`  — already in enum (line 51), semantically my "AttnVOut"
                 ("post softmax(Q@Kᵀ)@v, pre O-proj")

Only TWO are truly missing:

- `AttnScores`   — Q·Kᵀ / sqrt(head_dim), pre-softmax
- `AttnSoftmax`  — softmax(scores + causal_mask), pre-V

## Why it happened

Per `feedback_no_guessing.md`: should have run
`pmat query SaveTensorStage` BEFORE authoring v1.0.0. Instead I
extrapolated from the parent contract description without reading
the live enum source. Toyota Way andon — caught on next iteration.

Per `feedback_toyota_way_all_defects.md`: all defects are mine.
Fixing at the contract level BEFORE any implementation PR depends
on the wrong scope is exactly the cost-of-defect minimization
the toolchain is designed for.

## What v1.1.0 does

- Bumps version 1.0.0 → 1.1.0 PROPOSED (still pre-FUNCTIONAL)
- Reduces "new variants" from 5 to 2: AttnScores + AttnSoftmax
- Documents the FULL 9-stage layer-0 bisection chain spanning
  parent-contract stages + 2 new ones:

  attn_norm → qkv_matmul → qkv_bias → q_post_rope → k_post_rope
  → attn_scores [NEW] → attn_softmax [NEW] → attention → attn_out

- Updates all 5 falsifiers (SUB-001..005) to reflect reduced scope
- Adds bisection_chain_layer_0 equation pinning the 9-element
  cosine sequence (with empirical state per memory
  `2026-05-03 SHIP-007 finding`: cos[0]=0.99999995, cos[8]=0.9966)
- FALSIFY-ATTN-SUB-004 still BLOCKER_FIXTURE_ABSENT (pending HF
  FP16 oracle extension to capture 2 new stages on RTX 4090)

## Five Whys

1. **Why did v1.0.0 claim 5 new variants?**
   Authored without reading the live save_tensor_stage.rs source.

2. **Why didn't I read the source first?**
   Skipped the `pmat query SaveTensorStage` step that
   `feedback_no_guessing.md` mandates. Worked from the parent
   contract description's prose ("Embedding, AttnNorm, QkvMatmul,
   AttnOut, ...") which truncated 18 stages to 14.

3. **Why was the parent contract description truncated?**
   Doc-comment in `forward_traced_with_plan` rust source listed
   only 14 stages (the per-layer canonical-FFN order, omitting
   QkvBias + the parent's renamed Attention). My contract reused
   that prose instead of reading the enum directly.

4. **Why does this matter for SHIP-007 ship %?**
   It doesn't yet — the contract is still scaffold scope, no
   implementation PR has shipped against the wrong scope. v1.1.0
   correction lands BEFORE the cascade triggers.

5. **Why amend the contract instead of opening a sibling fix-PR?**
   Same branch (#1450) is the right place. Toyota Way: stop the
   line, fix the defect at source, then continue. A sibling PR
   would split the audit story across two commits with no benefit.

## Net effects

- Contract `trace-attn-sub-stages-v1` v1.0.0 → **v1.1.0 PROPOSED**
- `pv validate contracts/trace-attn-sub-stages-v1.yaml` exits 0
- MODEL-1 ship %: unchanged at 91% (this is contract correction)
- MODEL-2 ship %: unchanged at 57%
- Implementation cascade now correctly scoped to 2 new variants,
  not 5 — saves an estimated 60% of the enum-extension PR's LOC

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

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

* feat(aprender-serve): SaveTensorStage gains AttnScores + AttnSoftmax — FALSIFY-ATTN-SUB-001 PARTIAL_ALGORITHM_LEVEL

Implements `contracts/trace-attn-sub-stages-v1.yaml` v1.1.0 (PROPOSED, in PR #1450).

Adds the 2 new attention sub-stage variants to `SaveTensorStage`:

- `AttnScores`  — Q·Kᵀ / sqrt(head_dim), pre-softmax + pre-causal-mask
- `AttnSoftmax` — softmax(scores + causal_mask), pre-V-multiply

Closes the SHIP-007 layer-0 attention bisection gap inside the
Q·Kᵀ → softmax → ·V chain. The 9-stage layer-0 capture chain is now:

  attn_norm → qkv_matmul → qkv_bias → q_post_rope → k_post_rope
  → attn_scores [NEW] → attn_softmax [NEW] → attention → attn_out

## What changed

| File | Change |
|---|---|
| `save_tensor_stage.rs` | enum: 18 → **20** variants; `ALL` const, `canonical_name`, `FromStr` updated; doc-comment lists 21 names (incl. `layer_output` alias) |
| `save_tensor_stage.rs::tests` | Renamed `all_eighteen_*` → `all_twenty_*`; updated `is_per_layer_count` (18+2 = 20) + `canonical_names_match_contract_enumeration` to include the 2 new names; **4 new tests** for FALSIFY-ATTN-SUB-001 (round-trip, ordering, parser-list) |
| `save_tensor_plan.rs` | `all_keyword_expands_to_eighteen_stages` → `all_keyword_expands_to_twenty_stages`; `all_keyword_case_insensitive` count updated 18 → 20 |

## Test results

- `cargo test -p aprender-serve --lib inference_trace` — **167 passed, 0 failed**
- 4 new tests: `falsify_attn_sub_001_attn_scores_round_trip`,
  `falsify_attn_sub_001_attn_softmax_round_trip`,
  `falsify_attn_sub_001_2_new_stages_in_canonical_order`,
  `falsify_attn_sub_001_parse_list_accepts_2_new_stages_together`,
  `falsify_attn_sub_001_parse_list_accepts_full_attn_block_chain`
- `cargo check --workspace --lib` — clean

## Falsifier discharge

| ID | Status before | Status after | Why |
|---|---|---|---|
| FALSIFY-ATTN-SUB-001 | PARTIAL_ALGORITHM_LEVEL | **FUNCTIONAL** (eligible) | enum has 20 variants, parse_list accepts the 2 new tokens, ordering test passes |
| FALSIFY-ATTN-SUB-005 (additive purity) | PARTIAL_ALGORITHM_LEVEL | (no change yet — depends on `forward_traced_with_plan` threading, follow-up PR) |

Functional discharge of FALSIFY-ATTN-SUB-001 will be promoted in
`contracts/trace-attn-sub-stages-v1.yaml` v1.1.0 → v1.2.0 once this
PR + #1450 land. Today it stays PARTIAL_ALGORITHM_LEVEL because the
contract is still PROPOSED upstream.

## Five Whys

1. **Why this PR before #1450 lands?** Contract+impl can land
   together — #1450 introduces the contract, this PR provides the
   first implementation evidence. They reference each other and merge
   in either order without conflict.

2. **Why only the enum + tests, not `forward_traced_with_plan`?**
   Enum extension is the smallest atomic ticket per Toyota Way (one
   mechanism per PR). Threading the new variants through forward
   capture is the next PR (FALSIFY-ATTN-SUB-002 discharge).

3. **Why insert AttnScores+AttnSoftmax between KPostRope and Attention
   in `ALL`?** That's the canonical computation order pinned by the
   contract's ordering proof_obligation: `QkvBias → QPostRope →
   KPostRope → AttnScores → AttnSoftmax → Attention → AttnOut`.

4. **Why bump `ALL` count from 18 to 20 (not 19) when only 1 alias
   exists?** `LayerOutput` is a parse-only alias for `PostFfnResidual`,
   not a separate variant. The enum has 20 distinct variants; `ALL`
   excludes the alias only at the `FromStr` layer.

5. **Why include the 9-stage `parse_list_accepts_full_attn_block_chain`
   test?** The contract's `bisection_chain_layer_0` equation pins the
   9-element cosine sequence as the gate for FALSIFY-ATTN-SUB-004.
   This test pins the parser side of that gate so a future drift in
   stage names breaks loudly.

## Net effects

- 2 new `SaveTensorStage` variants land
- 5 new tests pin the variants + ordering + parser
- MODEL-1 ship %: unchanged at 91% (this is part of the SHIP-007
  bisection cascade; ship % moves when a falsifier flips DISCHARGED)
- MODEL-2 ship %: unchanged at 57%
- Implementation cascade ready to thread variants through
  `forward_traced_with_plan` next (FALSIFY-ATTN-SUB-002)

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

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

* feat(aprender-serve): wire 4 attention sub-stages in forward_traced_with_plan — FALSIFY-ATTN-SUB-002 PARTIAL_ALGORITHM_LEVEL

Stacked on #1451 (which adds the 2 new SaveTensorStage variants). When #1451
merges to main, this PR rebases cleanly and lands as a 4-stage wire fix.

## What this PR wires

| Stage | Existed in enum? | emit() existed? | After this PR |
|---|---|---|---|
| QPostRope   | YES | NO  | YES (new emit) |
| KPostRope   | YES | NO  | YES (new emit) |
| AttnScores  | NEW (#1451) | NO  | YES (new emit + accumulator) |
| AttnSoftmax | NEW (#1451) | NO  | YES (new emit + accumulator) |

Closes the parent-contract drift discovered in PR #1452 research evidence:
QPostRope + KPostRope were in the SaveTensorStage enum but had no emit()
calls in forward_traced_with_plan. The parent contract
`apr-cli-trace-save-tensor-v1.yaml` v1.4.0 (FUNCTIONAL) silently overstated
coverage for those 2 stages. This PR closes the drift as a side-effect.

## Implementation details

**QPostRope/KPostRope** (post line 133): emit q_all/k_all directly after the
inner loop populates them. Tensors already exist; this is just 2 emit()
calls — zero new allocation.

**AttnScores/AttnSoftmax** (inside head loop): allocate accumulator tensors
of shape `[num_heads × seq × seq]` ONLY when the plan requests them. Inside
the inner softmax loop, populate per (head, i, j) — zero overhead when
plan is None or doesn't ask for these stages (FALSIFY-ATTN-SUB-005:
additive purity).

Memory cost: BOS forward (seq=1) → num_heads * 1 * 1 * 4 bytes = 112 bytes
for Qwen2.5-Coder-7B (28 heads). Negligible. For longer seq, allocation
scales O(num_heads * seq^2) and is gated by plan.

## Test results

- `cargo test -p aprender-serve --lib -- --skip "gpu::"` — **13944 passed,
  0 failed, 51 ignored**
- `cargo check -p aprender-serve --lib` — clean
- inference_trace tests: 167/167 PASS
- (gpu:: tests have a pre-existing SIGABRT flake unrelated to this change)

## Falsifier discharge map

| ID | Status before | Status after | Why |
|---|---|---|---|
| FALSIFY-ATTN-SUB-002 (forward threading) | PARTIAL_ALGORITHM_LEVEL | (eligible for FUNCTIONAL once contract YAML on main + this lands) | 4 emit() calls now thread the 4 stages in canonical order |
| FALSIFY-ATTN-SUB-005 (additive purity) | PARTIAL_ALGORITHM_LEVEL | (eligible) | accumulator allocation gated by plan.should_save() |

## Five Whys

1. **Why wire 4 stages, not 2?** QPostRope + KPostRope are pre-existing
   gaps in the parent contract; the same-file fix is a free side-effect
   per Toyota Way "all defects are mine".

2. **Why allocate accumulators only when requested?** O(num_heads * seq^2)
   memory shouldn't be paid on the default forward path. Plan-gating
   keeps the production inference path zero-overhead.

3. **Why insert capture at lines 133, 152, 160 specifically?**
   Per `evidence/ship-007-layer0-attn-bisection-2026-05-04/forward-traced-research.md`:
   line 133 = post Q/K/V copy (Q/K post-rope), line 152 = scores after
   scale (pre-softmax), line 160 = post-softmax probs.

4. **Why use scores_all.is_some() check vs always-allocate?**
   Always-allocate forces O(seq^2 * num_heads * 4) bytes per layer
   regardless of capture. Some(Vec) idiom plus is_some_and check is the
   idiomatic Rust pattern for conditional capture.

5. **Why this PR stacked on #1451 rather than off main?**
   Requires SaveTensorStage::AttnScores + AttnSoftmax variants, which only
   exist on #1451's branch. When #1451 merges, this rebases to main as a
   clean 51-line delta.

## Net effects

- 4 stages now wired in `forward_traced_with_plan`
- MODEL-1 ship %: unchanged at 91% (stays scaffold; ship % moves at
  FALSIFY-ATTN-SUB-004 LIVE DISCHARGE in a future cycle)
- MODEL-2 ship %: unchanged at 57%
- Cascade step 4/8 of §47.1 roadmap delivered

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

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

---------

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

…tion cascade ALGORITHM-LEVEL COMPLETE (#1458)

After §47 recorded the cascade-started milestone (PRs #1450 + #1451 + #1452
scaffolding), the same-day continuation cycle closed §47.1 cascade roadmap
steps 4-6 at the algorithm level via PRs #1455, #1456, #1457.

## What landed (§47.1 cascade roadmap)

| Step | PR | Discharge |
|------|----|-----------|
| 4 | #1455 | FALSIFY-ATTN-SUB-002 PARTIAL_ALGORITHM_LEVEL — wires `QPostRope`+`KPostRope`+`AttnScores`+`AttnSoftmax` in `forward_traced_with_plan`; closes §47.4 parent-contract drift as side effect |
| 5 | #1456 | FALSIFY-ATTN-SUB-003 algorithm-level pinned via 2 drift-prevention tests; 0 LOC production change (loader is genuinely per-stage-agnostic, as spec predicted) |
| 6 | #1457 | FALSIFY-ATTN-SUB-004 BLOCKER_FIXTURE_ABSENT → PARTIAL_ALGORITHM_LEVEL on merge — extends `scripts/generate_qwen25_coder_fp16_stages.py` with `--with-attn-substages` (default ON) installing per-instance `Qwen2Attention.forward` monkeypatch under `attn_implementation="eager"` |

## Toyota Way correction during research (PR #1457)

The pre-impl research note estimated **7 missing stages, ~140 LOC**. Live source inspection during PR #1457 found **3 already captured** via existing forward hooks (`make_qkv_hook` derives qkv_matmul/qkv_bias from q_proj/k_proj/v_proj outputs via bias subtraction; `hook_o_proj_pre` captures `attention` as input to o_proj). Net: **4 stages, ~80 LOC monkeypatch**.

Per `feedback_no_guessing.md`. Cost-of-defect paid at the implementation layer (cheapest place once the research note had been authored from outdated docstring lines).

## Steps 7-8 require operator action

| Step | Blocker | Workaround |
|------|---------|-----------|
| 7 LIVE | (a) canonical `apr` binary built pre-#1451 — rejects `attn_scores` stage. (b) PyTorch/CUDA driver mismatch on host. | (a) `cargo build --release --features cuda --bin apr`. (b) operator updates driver OR `--device cpu` (multi-min). |
| 8 fix | Gated on step 7 bisection finding. | n/a — discovery-driven scope. |

## Net effects

- Spec v2.92.0 → **v2.93.0**.
- §47.1 cascade roadmap: **6/8 steps algorithm-level COMPLETE**; steps 7-8 LIVE/operator-gated.
- Coverage tally: 20+32 → **20+36** (+4 PARTIAL_ALGORITHM_LEVEL from `trace-attn-sub-stages-v1` v1.1.0 falsifiers landing on main when #1450 merged: SUB-001/002/003/005). SUB-004 stays BLOCKER until #1457 ships.
- **MODEL-1 ship %**: unchanged at **91%** (cascade is scaffold; ship % moves at SUB-004 LIVE DISCHARGE in step 7).
- **MODEL-2 ship %**: unchanged at **57%**.

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

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