Fix README images and links on PyPI (0.3.1) (#1)

## What

- **PyPI README fix:** PyPI can't resolve the README's relative paths,
so the logo, both demo GIFs and every doc link are broken on
https://pypi.org/project/trainjudge/. `hatch-fancy-pypi-readme` now
rewrites relative images to `raw.githubusercontent.com` and relative
links to `github.com/.../blob/main/`, **only in the copy sent to PyPI**.
The README on GitHub is unchanged.
- **`AGENTS.md`:** tells coding agents that `main` is protected (work on
a branch, open a PR, CI must pass).
- **Version 0.3.1** in `pyproject.toml`, `__init__.py`, `plugin.json`,
`gemini-extension.json`, the README badge and the CHANGELOG.

## Why

The repo just went public and the launch post goes out today. The PyPI
page is the first thing `pip install` users see.

## How it was tested

- Built the wheel locally and parsed its PyPI description: **0 relative
links left**.
- Checked every rewritten URL with curl: the logo (light and dark), both
demo GIFs and the doc links all return **200**.
- `pytest`: 351 passed, 3 skipped; `ruff check` is clean.

## After merging

Create the release with `gh release create v0.3.1`. That publishes to
PyPI through Trusted Publishing, and the fixed README shows up there.

Co-authored-by: Himanshu <himanshu@Himanshus-MacBook-Air.local>

Add label and custom-scorer evals, Cursor and Gemini integrations

- Label eval: classification tasks are detected from the test split and
  scored by label accuracy, with macro-F1, per-label F1 and top confusions.
- Custom scorers: eval/verify --scorer path.py:func (or module:func) score
  any task with a user function. The scorer is recorded in the run and its
  hash invalidates saved evals when it changes. Ships an example,
  trainjudge.scorers.key_facts, for prose answers.
- Intent routing demo for the label task (customer support).
- Cursor rule and Gemini CLI extension; AGENTS.md and the skill install
  from PyPI and explain scorers; a test keeps them in sync.
- Model cards for PyTorch runs name peft and show transformers usage.
- Version 0.3.0.

TrainJudge v0.2.1: GPU memory fix, internal cleanup, mypy in CI. Firs…

…t PyPI release.

TrainJudge v0.2.0

Train and verify on NVIDIA GPUs (PyTorch backend, verified on a T4) as
well as Apple Silicon (MLX), with a Colab notebook for anyone without a
GPU. Domain packs and demos for healthcare, legal, e-commerce, customer
support, HR, BFSI and education. JSON extraction eval, PII masking,
region-aware sensitive-data detection, a baseline cache and milestone
progress for agents. Relicensed to Apache-2.0. See CHANGELOG.md.

TrainJudge v0.1.0

Diagnose whether fine-tuning is the right fix, train locally with MLX LoRA,
and verify the result on held-out task metrics and a general-capability
regression suite (IMPROVED / REGRESSED / REJECTED).