…cision (#1449)

After §45 landed the 5/5 DISCHARGE milestone for `apr-cpu-vs-gpu-output-parity-v1`,
the natural follow-up question is whether the 238 commits accumulated since v0.31.2
(2026-04-19) warrant a `cargo publish` cut today.

## Verdict: HOLD

The release-readiness audit found exactly one load-bearing blocker — SHIP-007
layer-0 attention divergence is empirically pinpointed (cos=0.99999995 attn_norm
→ 0.9966 attn_out per memory `2026-05-03 SHIP-007 finding`) but **not yet fixed**,
so cutting v0.32.0 today would crates.io-ship a binary where `apr run` on a 7B
GPU teacher still emits gibberish unless the user passes `--no-gpu`.

The §41-§45 jidoka armor makes the failure visible + fail-closed (which is
shippable behaviour), but a user-facing `## [0.32.0]` headline that reads
"5/5 DISCHARGE on apr-cpu-vs-gpu-output-parity-v1" implies the GPU correctness
hole is closed when in truth it is only contained.

Per `feedback_fix_root_cause_never_route_around.md`: route-around-via-fallback
is acceptable as a *temporary jidoka layer*, but it is muda to ship a release
whose headline claims a fix that doesn't exist.

## What §46 records

| Subsection | Content |
|---|---|
| 46.1 | What's accumulated since v0.31.2 (8-row headline table) |
| 46.2 | Release-readiness gate audit (6-gate verdict table) |
| 46.3 | Why SHIP-007 is the load-bearing blocker |
| 46.4 | Pre-flight artifacts shipped alongside this decision (PR #1448) |
| 46.5 | **Pre-conditions for the v0.32.0 cut** — 5 falsifiable gates |
| 46.6 | Five Whys (why hold / why now / why §46 not extend §45) |
| 46.7 | Net effects + ranked open follow-ups |
| 46.8 | Spec amendment cadence preserved |

## Pre-conditions (§46.5) for the future v0.32.0 cut

1. SHIP-007 layer-0 attention divergence FIXED (cos ≥ 0.999 at every sub-stage)
2. PR #1448 merged (CHANGELOG `[Unreleased]` populated; README drift gate GREEN)
3. Workspace version bumped 0.31.2 → 0.32.0; `## [0.32.0] - <date>` heading
4. Post-publish QA: `cargo install aprender --force` + /dogfood GO verdict
   (per `feedback_post_publish_qa_required.md` — v0.31.1 was yanked for skipping)
5. Drift gates GREEN: `check_readme_claims.sh` + `pv validate` + `cargo deny`

## Net effects

- Spec v2.90.0 → **v2.91.0**
- MODEL-1 ship %: unchanged at 91% (this is metadata, not a falsifier flip)
- MODEL-2 ship %: unchanged at 57%
- Coverage tally: unchanged (no PARTIAL → DISCHARGED in this cycle)
- Single highest-leverage next-session deliverable: SHIP-007 layer-0 attention
  bisection using `apr trace --save-tensor` + HF FP16 oracle from PR #1423

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

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

…pre-existing capture gaps (QPostRope + KPostRope) (#1452)

Records pre-implementation research for FALSIFY-ATTN-SUB-002 (`trace-attn-sub-stages-v1.yaml` v1.1.0).

## What this evidence pins

While researching where to wire AttnScores + AttnSoftmax in
`forward_traced_with_plan` (per the v1.1.0 contract), discovered that
QPostRope + KPostRope variants exist in the SaveTensorStage enum
(lines 47-50) but have **no `emit()` call** in `forward_traced_with_plan`.

The RoPE-rotated tensors q_all + k_all are computed at lines 130-131
but never captured. The parent contract `apr-cli-trace-save-tensor-v1.yaml`
v1.4.0 (FUNCTIONAL) silently overstates coverage for these 2 stages.

## What FALSIFY-ATTN-SUB-002 will wire

When #1451 lands, the next PR will wire 4 capture points (not 2):

| Stage | Source line | Existed in enum? |
|---|---|---|
| QPostRope   | post line 133 | YES (gap) |
| KPostRope   | post line 133 | YES (gap) |
| AttnScores  | line 152 (per head, accumulator) | NEW (#1451) |
| AttnSoftmax | line 160 (per head, accumulator) | NEW (#1451) |

## Why an evidence file, not a 5th stacked PR

Four PRs (#1448-#1451) already in flight. A 5th stacked PR would
slow CI throughput. Recording the implementation plan here so the
next loop iteration can spawn the impl PR off main once #1451 merges.

## Five Whys + cross-references

In `evidence/ship-007-layer0-attn-bisection-2026-05-04/forward-traced-research.md`:

- Five Whys for scope (4 stages, not 2)
- Wire-plan with insertion points
- Backward-compat test plan
- Next-iteration deliverables checklist

## Net effects

- Evidence file lands; no code change in this PR
- MODEL-1 ship %: unchanged at 91%
- MODEL-2 ship %: unchanged at 57%
- Unblocks the next loop iteration's atomic PR

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

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