task: drop unused numpy from the ctl and all extras

[dgarros]

Component

Python SDK, infrahubctl CLI

Task Description

numpy is declared in both the ctl and all optional-dependency groups, but the SDK never imports it. A repo-wide code search for numpy matches only pyproject.toml — there is no import numpy anywhere in the source.

It looks vestigial from when pyarrow hard-depended on numpy. That is no longer the SDK's problem to declare:

pyarrow	own requires_dist
14.0.0	numpy>=1.16.6
16.0.0	numpy>=1.16.6
20.0.0	(no numpy)
25.0.0	(none at all)

So removing the line is safe in both directions: at the pyarrow>=14 floor, pyarrow pulls its own numpy transitively; on modern pyarrow, numpy simply isn't needed. Either way the SDK is restating a dependency it doesn't own and doesn't use.

Why it's worth removing

It drags ~40 MB of BLAS-backed wheels into every ctl/all install for a package nothing dereferences.

The dual markers force resolver forking in every downstream lockfile:

"numpy>=1.24.2; python_version<'3.12'",
"numpy>=1.26.2; python_version>='3.12'",

It actively obstructs Python 3.14 adoption. The SDK already advertises requires-python = ">=3.10,<3.15", so 3.14 is in scope. But numpy 1.26.4 — a very common resolution for the >=1.26.2 floor — publishes wheels for cp39–cp312 only (no cp313, no cp314) while declaring requires_python: >=3.9 with no upper bound. Resolvers therefore consider it eligible on 3.14 and fall back to a source build that doesn't succeed. The first numpy with cp314 wheels is 2.3.2.

A fresh resolve dodges this by picking numpy 2.5.x. An existing lockfile does not: downstream, widening requires-python to admit 3.14 left numpy pinned at 1.26.4 and broke uv sync --python 3.14. We had to bump numpy to 2.4.6/2.5.1 purely to unblock a package that nothing in the dependency tree imports.

Suggested change

Drop these two lines from both the ctl and all groups in pyproject.toml:

 ctl = [
     ...
-    "numpy>=1.24.2; python_version<'3.12'",
-    "numpy>=1.26.2; python_version>='3.12'",
     "pyarrow>=14",

Optionally raise the pyarrow floor past the version that dropped numpy, if avoiding numpy entirely (rather than just not declaring it) is the goal.

Related

#1160 (fix(ctl): import pyarrow lazily in the JSON importer) trimmed this same corner for slim installs. Dropping the unused numpy declaration is the natural follow-up — it's the last thing making ctl/all heavier than the code requires.

Verified against

• infrahub-sdk 1.22.0 (installed), and pyproject.toml on develop, infrahub-develop, and stable — the declaration is present and unchanged on all three.