chore(register): C-62 — the pinned pipeline-core release still installs 2.5 GB of CUDA

[Polichinel]

Answers the maintainer's question during the development→main sweep: "Is geopandas back? Is it still here?"

Short answer

Not in the code. Yes in the environment. And the real cost isn't geopandas.

Source level — clean

Three mentions of geopandas/shapely in .py/.toml, all asserting its absence:

enrichment.py:9          "No geopandas, no shapefiles, no spatial computation."
build_gaul_lookup.py:11  "...only a table join and a coordinate formula. No geopandas."
test_doc_accuracy.py:29  a regex that BANS the word from docs

C-39's deletion held completely.

Environment level — not clean

poetry.lock resolves geopandas 1.0.1, optional = false. The carrier is the pinned release:

views-pipeline-core 2.3.0 (PyPI) declares:
  torch >=2.6.0,<3.0.0 | geopandas >=1.0.1,<2.0.0
  scipy | seaborn | plotly | plotly-express

Their development already dropped geopandas and their own falsification tests name the problem. It's gated purely on publishing 3.0.0.

Measured, and it inverts the premise

Live project venv — 2.8 GB, none of it imported here:

Installed, never imported	Size
nvidia/ (torch CUDA runtime)	2.5 GB
plotly	42 MB
matplotlib	25 MB
geopandas	1.6 MB

The audit started as "is geopandas back?" — and geopandas is the architectural violation, the thing PR #42 spent a 3,171-line deletion removing. But it's 1.6 MB. The material cost is torch's CUDA stack at 89% of the venv, in a repo with no GPU code, no training, no tensor operations.

Optimising for the offensive dependency rather than the expensive one would have missed nearly the whole bill. That correction is recorded in the entry, not just the commit.

Why C-62 and not an update to C-44

They pull in opposite directions, deliberately:

• C-44 holds the 3.0.0 bump — the maintainer's standing constraint that the platform must run smoothly on development across all repos first. Still right.
• C-62 records what waiting costs.

Folding this into the entry arguing for the delay would hide the bill inside the argument. Tier 3 — installed but never imported, so no correctness impact; the cost is footprint and a first release (#125) that sets expectations.

Cross-repo, filed today

• views-datafactory#387 — corrected. My original text said the equal-area check "needs geopandas." Wrong: that repo already does spatial work geopandas-free (shapely STRtree + pyproj, ADR-030 compliance). Left uncorrected it would have invited reintroducing it.
• views-pipeline-core#319 — consumer data point with the measured numbers. Not a request to accelerate.

Verification

Register integrity guard green. 62 concerns, 25 open. Cluster G. No source code touched.

🤖 Generated with Claude Code