paper: pass-2 consistency — purge version labels from App A/B/bib

[FluffyAIcode]

Reviewer feedback addressed

"论文的'评测协议'和'版本边界'在全文内部自相矛盾，影响可信度"

• 摘要和正文把这篇稿子写成一个统一的 KakeyaLattice family paper，同时包含 D4 和 E8 两个变体
• 正文主结果 §5 是 D4 + snapshot-mode + 4 passages；§6 的 E8 又切换成 in-forward + no-boundary + n=32 + 95% CI
• §7 caveat 明确写了："All numbers in this paper are snapshot-mode" — 与 §6 直接冲突
• 附录的复现实验混入了 v1.5 rigorous evaluation harness；附录 B/C 又把 D4 写成 v1.4、E8 写成 v1.5

What was already fixed in prior PRs

The protocol self-contradiction in §7 caveat and the mixed protocol framing across §5/§6 were already fixed in PR #41 (paper: fix protocol self-contradiction and bit-accounting inconsistencies, merged to main as 9493fac). At main HEAD on 2026-04-24:

• §1 has a "Strict-GPU, live-vLLM; two measurement protocols" paragraph declaring both protocols up-front.
• §4 has a new §4.4 "In-forward rigorous evaluation" subsection defining n=32 / 95% CI.
• §5 title: Main results: D_4 variant, snapshot protocol. §5 intro: All tables in this section use the snapshot protocol defined in §4.2 (n=4 passages, boundary-layer protection k=2). The protocol isolates codec intrinsic error...
• §6 preamble: Every result in this section uses the in-forward rigorous evaluation protocol defined in §4.4.
• §7.2 caveat no longer says "all numbers are snapshot-mode"; instead maps each protocol to its question: Snapshot-mode isolates codec intrinsic error and is used for every D_4 result in §5... In-forward rigorous measures deployment-realistic error and is used for every result in §6.
• All results-table captions cite their protocol tag.

What this PR (Pass 2) fixes — the residual version-label leakage

Reviewer correctly noted that even after PR #41, the appendices and bibliography still leaked version labels and contradicted the paper's own framing choice (lattice-name-only in body). Specifically:

App A (Reproducibility manifest)

• Opening line said introduced with the v1.5 release, framing the harness as a v1.5-era artifact instead of a codec-agnostic, protocol-configurable tool.
• Code-block comment said # PPL / MSE / CR (v1.4 + v1.5 + TurboQuant at matched Q / b), treating v1.4 / v1.5 as codec identities.
• Directory paths reports/v1_5_release/ / reports/v1_4_release/ referenced without explaining that these are on-disk git-tag directory names.

Fixed: rewrote opening paragraph (both lattice variants), code-block comment (# In-forward rigorous (n=32, 95% CI): D4, E8, and TurboQuant, same run), --out-dir defaults (now reports/rigorous_eval/, protocol-generic). Added closing paragraph explaining that --v15-q-values is a CLI-flag identifier and reports/v1_{4,5}_release/ are git-tag-matched directory paths, not codec names.

App B (was "Canonical naming", now "Implementation identifiers")

• The appendix was called "Canonical naming" and used v1.4 codec (D_4) / v1.5 codec (E_8) as the canonical names, with "spoken / written: v1.4 kakeya zamir lattice GPU" entries. This directly contradicted the body prose from PR paper: remove version labels from title/body, present as D4/E8 lattice variants #40 which references codec variants exclusively as "$D_4$ nested lattice" / "$E_8$ nested lattice".

Fixed: Renamed to "Implementation identifiers". First sentence: The paper is agnostic to version labelling: the two codec variants are the D_4 nested lattice and the E_8 nested lattice, and every result table cites its protocol. This appendix lists only the repository-level identifiers needed to reproduce the bit-identical codec output. Removed all "spoken / written" entries. Final disclaimer: The v1.4 / v1.5 strings appear only as git release tags and as the V14/V15 class-name prefixes chosen by the repository authors; the paper itself references the codec variants exclusively by their lattice names.

Bibliography (kakeya-v14-release, kakeya-v13-paper)

• KakeyaLattice v1.4: the canonical implementation → KakeyaLattice: the canonical implementation (drop version suffix in title).
• Package reference kakeyaturbo_py.V14KakeyaZamirLatticeGPU → Python package kakeyalattice (classes V14KakeyaZamirLatticeGPU and V15KakeyaZamirE8GPU). kakeyaturbo_py was renamed to kakeyalattice in PR paper: remove version labels from title/body, present as D4/E8 lattice variants #40; this was a regression.
• Harness path benchmarks/multimodel_v14_vs_tq.py → benchmarks/rigorous_eval.py (the actual current canonical harness).
• Snapshot-hook path vllm_backend/kakeya_v1_3_ppl/snapshot_hook.py → vllm_backend/kakeya_v1_4_snapshot/snapshot_hook.py (matches the post-rename repository layout).
• Superseded by this paper at tag v1.4. → Prior unpublished draft, superseded by this paper. (removed version-label leakage from the prior-work reference).

What remains literal v1.4 / v1.5 — intentional, disclaimed

• Bib entry kakeya-v14-release: lists Release tags: v1.4 (first D_4 release, commit 6b02711) and v1.5 (first E_8 release). These ARE the tags the bib points at.
• App A closing paragraph: explicitly labels --v15-q-values and reports/v1_{4,5}_release/ as repository-level identifiers, disclaiming their role as codec identities.
• App B: class names V14KakeyaZamirLatticeGPU / V15KakeyaZamirE8GPU; module names kakeyalattice.v1_4_kakeya_zamir_lattice_gpu / kakeyalattice.v1_5_kakeya_zamir_e8_gpu; release tags. These are the real Python symbols and git refs a reader needs to reproduce bit-identical codec output. The closing disclaimer in App B explicitly notes that these are repository artifacts.

Post-fix audit

$ rg -n "v1\.4|v1\.5" reports/paper/kakeyalattice.tex
1358:  Release tags: \texttt{v1.4} (first $D_4$ release, commit
1359:  \texttt{6b02711}) and \texttt{v1.5} (first $E_8$ release).
1519:$E_8$-variant codec registered at release tag \texttt{v1.5};
1539:    release tag \texttt{v1.4} (commit \texttt{6b02711}).
1543:    release tag \texttt{v1.5}.
1546:The \texttt{v1.4} / \texttt{v1.5} strings appear only as git release

Every remaining v1.4 / v1.5 is either a bib entry release tag, a class/module reproducibility identifier, or the explicit disclaimer that the paper uses lattice names only. Zero body-prose or table-cell leakage.

Build

$ cd reports/paper
$ rm -f *.aux *.log *.out *.toc *.fls *.fdb_latexmk *.pdf
$ latexmk -pdf -interaction=nonstopmode -halt-on-error kakeyalattice.tex
...
Output written on kakeyalattice.pdf (15 pages, 452079 bytes).

15 pages (unchanged from PR #42 trim), 452 KB (+1 KB from new App A explanatory paragraph), zero undefined references, zero LaTeX errors.

Files

• reports/paper/kakeyalattice.tex (+55 −39 lines)
• reports/paper/kakeyalattice.pdf (regenerated)

No code / hook / codec / benchmark / release-artifact changes in this PR.

Not for auto-merge

Per user's "先不要合入" pattern from the Stage 0.5 PR #43 thread, this PR is filed as draft. Review the App A paragraph + App B structure; merge when satisfied with the reproducibility-vs-brand-consistency trade-off.