results+paper(z-gap): Strategy D 7-model results — 35/35 cells, P1 partial via E5 family

[heznpc]

Strategy D extension ran successfully after einops install:

Model en ko zh ar es agg
UniXcoder 1.22* 1.01* 1.08* 1.01* 1.05* 1.07
MiniLM-L12 1.23* 1.12* 1.18* 1.10* 1.19* 1.16
Nomic v1.5 1.24* 1.02* 1.03* 1.01* 1.07* 1.07
E5-small [NEW] 1.22* 1.09* 1.13* 1.09* 1.14* 1.13
E5-base [NEW] 1.22* 1.11* 1.13* 1.11* 1.16* 1.14
E5-large 1.28* 1.16* 1.19* 1.16* 1.22* 1.20
BGE-M3 [NEW] 1.21* 1.14* 1.16* 1.14* 1.16* 1.16

35/35 cells: R_code > 1, p < 0.05 after Holm-Bonferroni
Permutation null mean: R in [1.000, 1.005] across all cells (C2 baseline confirmed)

Findings:

• Cross-model robustness: alignment holds across code-trained (UniXcoder,
  Nomic), hybrid (BGE-M3), and NL-only (MiniLM, E5 family) architectures.
• E5 family scale-convergence (within same architecture/training recipe):
  1.13 (384d) -> 1.14 (768d) -> 1.20 (1024d). Partial P1 support — monotonic
  qualitatively but non-linear (small->base flat, base->large steep).
• D_train modulation re-confirmed at scale: English R_code 1.21-1.28,
  Korean/Arabic 1.01-1.16, tracking code-corpus language representation.

Paper:

• §5.5 Table: 4 rows -> 7 rows. Caption now reports 35/35 cells + null R
  range. Body text rewritten from "20/20" to "35/35" + new "Third pattern"
  paragraph on E5-family partial scale-convergence.

Dependencies:

• einops>=0.7 added to experiments/requirements.txt and pyproject.toml.
  Required by nomic-ai/nomic-embed-text-v1.5 trust_remote_code module; the
  M3 try/except wrap correctly isolated the first-run failure, allowing
  one targeted dep fix instead of a wholesale debug.

Decisions log:

• planning/decisions.md: 2026-05-21 entry documenting einops fix and the
  35/35 result.