…hesis — spec v2.74.0 → v2.75.0

Live diagnostic on noah-Lambda-Vector RTX 4090 against canonical 7B teacher
(/mnt/nvme-raid0/models/ship-two-001/qwen2.5-coder-7b-instruct-q4k.apr) shows
that §28's mechanical "replace helpers::f32_matmul with Q4K-fused dispatch"
fix would NOT close the 9× layer-0 qkv std gap that propagates to layer 3's
18.23× ffn_swigl ratio.

Two empirical findings refute the narrow hypothesis:

1. q4k_layers IS fully populated for all 28 layers (Q=7,225,344 bytes per
   layer, K=V=1,032,192 each, gate=up=down=38,191,104 each) —
   §28.4 option (a) is already shipped.

2. APR's F32 fused-qkv weight (load_qkv_weight at mod_dequant_q4k_apr.rs:181)
   is numerically equivalent to per-Q/K/V Q4K dispatch within Q4K rounding:
   max |diff|=0.005294, RMS diff=0.000673 on layer-0 Q-projection. The
   matmul kernel switch in §28 would change <0.5% of std.

The 9× layer-0 qkv divergence (APR=10.33 vs GGUF=1.14) is REAL but lives
upstream of the gate matmul. Three candidate sites for §30.4 bisection:

- qkv_bias add at pmat-260.rs:332-334 (mean shift APR=0.2559 vs GGUF=-0.0163
  is suggestive)
- RoPE precision at pmat-260.rs:377-378 apply_rope_f32
- Per-head Q/K RMSNorm at pmat-260.rs:359-374 (should be skipped for
  Qwen2.5-7B but worth verifying)

PR E paused. Coverage scoreboard unchanged at 15+33 (§29 §30.5).

Files added:
- crates/aprender-serve/examples/check_q4k_population.rs (q4k_layers dump)
- crates/aprender-serve/examples/diag_apr_qkv_layer0.rs (F32 vs Q4K compare)
- evidence/ship-007-pr-e-investigation-2026-04-27/findings.md (full analysis)
- evidence/ship-007-pr-e-investigation-2026-04-27/check_q4k_population.txt
- evidence/ship-007-pr-e-investigation-2026-04-27/diag_apr_qkv_layer0.txt

Spec amendment:
- §30 — PR E investigation refutes §28 narrow hypothesis
  - 30.1 Diagnostic evidence
  - 30.2 What §28 got right and got wrong
  - 30.3 What's still load-bearing
  - 30.4 Falsifiable next investigation step
  - 30.5 Coverage scoreboard (unchanged)
  - 30.6 Methodology note — investigative falsification IS the discharge
- Header: v2.74.0 → v2.75.0

Methodology: per feedback_fix_root_cause_never_route_around.md, the §28
mechanical fix would have route-around'd a real bug because the named site
(matmul kernel) is not where the divergence originates. The empirical
refutation is itself a coverage-protecting artifact — it prevents a no-op
fix from shipping.

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