V2

[himjl]

Summary

This branch is major v2 overhaul of hobj. It modernizes packaging and distribution (now using PyPi and uv), replaces the old ad hoc data/image loading flow with packaged OSF-backed datasets and a local cache, and reshapes the public API around a smaller set of supported entry points for loading data, building learners, and running the two published benchmarks.

What Changed

• Moved the project from setup.py/requirements.txt to pyproject.toml + uv.lock, bumped the package to hobj==2.0.0, and standardized on Python 3.12.
• Added local developer tooling and release plumbing, including a Makefile, GitHub Actions CI, and a PyPI-oriented install/publish workflow.
• Introduced packaged dataset download and resolution via hobj.data.download, including safe archive extraction, versioned caching under the user home directory, and a new hobj-download-data CLI.
• Replaced the previous URL-oriented image loading path with manifest-driven local image access, including list_image_ids, get_image_path, load_image, and structured loaders for the high-variance, one-shot, warmup, and catch image sets.
• Replaced older behavior-loading code with CSV-backed loaders for the packaged high-variance and one-shot human datasets, with validation and support for custom cachedir roots.
• Populated the top-level hobj API so users can import the main benchmarks, data loaders, image loaders, create_linear_learner, and RandomGuesser directly from the package root.
• Refactored benchmark internals into clearer binary-classification and generalization modules, with session simulation, bootstrap summaries, variance-corrected MSEn scoring, and xarray-based statistics outputs.
• Simplified the learning-model surface around a binary learner abstraction plus a calibrated linear learner factory that works from precomputed image features.
• Removed or relocated older legacy code paths that are no longer part of the supported package surface, including the previous behavioral-data, image, statistics, config, and state-space scaffolding.
• Rewrote the README and examples, added notebooks, and documented post-publication renames/statistical changes in site/changelist.md.

Reviewer Notes

• This is a large branch: 163 commits and a broad package/layout refactor.
• Data access now assumes packaged OSF artifacts and local manifests instead of the earlier direct-download code paths.

Testing

• Added CI on Python 3.12.
• Added tests for cache/download resolution, packaged behavior/image loaders, simulation primitives, statistics builders, dataset-shaped benchmark outputs, and deterministic dummy-learner behavior.