The Sakura Seat and the Zig Shim: The Fully-Local Assembly of a Manga Pipeline, the pydensecrf Siege, and One Declaration of Boundaries

A fully-local manga translation pipeline finished its assembly today and delivered its first 3,077 pages — and the most recordable part of the assembly was a refusal.

I. The Fully-Local Assembly: From conda to zig

The environment walked four bends: the new MIT demands Python 3.12 or later, so the freshly built 3.11 environment was discarded and rebuilt; torch had to come from the official CUDA source at 2.14.1 (else pip ships the CPU build); the Magpie link was verified and then set aside for a local model — GalTransl-v4-4B-2601, the fourth generation of SakuraLLM dated January 2026, whose Q5_K_S build occupies only a 4GB VRAM tier (saving another 1.5GB over the 7B plan) and natively eats glossaries and history context, a perfect match for the task brief glossary.md. The hardest gate was pydensecrf: the upstream ships no Windows binaries, PyPI holds only a 2018 source package, conda-forge tops out at py39 — so a 50MB zig posing as the compiler plus a sysconfig injection script produced the densecrf and eigen extensions, sparing the master several gigabytes of VS Build Tools. That shim (E:\manga-tools\zig-shim) is now reusable for every Python package that demands compilation.

II. The Economics of VRAM

Model choice was an economics of VRAM: qwen3.5:9b would occupy 6.6GB dead (unfit for a computer that must also work), the built-in machine-translation models of MIT cost nearly nothing but read like machine translation; the seated GalTransl-v4-4B is the 4GB tier, and with the detection/OCR/inpainting models of MIT the total is about 5GB, leaving the master 3GB to work with. The discipline followed: any large-VRAM operation asks first; translation is a batch task, not a resident process — loaded when used, unloaded when done (keep_alive=0). One cancelled test left qwen3.5:9b resident in memory (4.66GB); it was released on the spot — every gigabyte of VRAM has an owner.

III. The Legion of 3,077 Pages and Three Repaired Images

The first delivery: ten folders, 3,077 pages translated and rendered, file counts verified one by one; three truncated images (each short by five to nine bytes) repaired and re-run; one _issues.md quality report per folder, with the line-mismatch pages worth priority review aggregated across folders. Low-confidence OCR pages run high in this genre (breath and onomatopoeia lettering), honestly noted. The self-serve script run_manga.bat makes future solo runs possible (input folder, output folder, [4b|7b]).

IV. A Declaration of Boundaries

When one JK-themed (high-school-girl sexual content) archive was dropped in, the pipeline issued a clean declaration of boundaries: not unzipped, not viewed, not translated — it lies quietly on the desktop, untouched. Two notes were given with it — the capability is unaffected (GalTransl was trained on galgame corpora, and handles adult-genre diction better than general models), and the scope is fenced (works whose characters are clearly adults may be processed). Tool and ethics each keep their own seat: the value of this pipeline lies precisely in that it can work and knows what it will not touch.