Cache Hugging Face tokenizers in CI

[xeophon]

Summary

• add a Hugging Face cache directory for the Test workflow
• restore/save that cache in the existing Python test matrix
• pass HF_TOKEN through to the main Verifiers test job to reduce public Hub rate-limit pressure

Context

This is the smaller follow-up for the Hugging Face 429 flake observed on #1529. It keeps the existing Test workflow topology and avoids the separate cache-warm job.

Validation

• ruby YAML parse for .github/workflows/test.yml
• git diff --check

---

Note

Low Risk
Workflow-only changes (cache path, secrets env, cache step) with no application runtime or auth logic changes.

Overview
The Test workflow now pins Hugging Face artifacts under a workspace-local HF_HOME and restores/saves that directory with actions/cache, keyed off uv.lock and the renderer test files that drive tokenizer downloads.

The main Verifiers matrix job also receives HF_TOKEN from repo secrets so Hub access is less likely to hit public rate limits (addressing CI 429 flakes on tokenizer loads).

^{Reviewed by Cursor Bugbot for commit 1811e61. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Cache Hugging Face tokenizers in CI to speed up test runs

Adds a caching step in test.yml that stores and restores the Hugging Face tokenizer directory (HF_HOME) across CI runs. The cache key is derived from the OS and hashes of uv.lock and the relevant test files, with a fallback restore key. HF_TOKEN is also now exposed to the job via repository secrets.

^{Macroscope summarized 1811e61.}

[macroscopeapp]

Approvability

Verdict: Approved

CI-only change that adds caching for Hugging Face tokenizers to speed up test runs. Standard GitHub Actions caching pattern with no impact on production code or runtime behavior.

^{You can customize Macroscope's approvability policy. Learn more.}