docs: document the dataset format, fix the vendor command, and name the PyPI collision

[TMHSDigital]

Fixes #29. Fixes #60. Fixes #61.

#61: the dataset format was undocumented. New page, docs/datasets.md ("Your own data"), hosted on the site under Using plumbline. It gives:

• the fields: id, text, labels, gold_label, and the optional question_type and label_descriptions;
• what each question type means, and an example row of each kind;
• what the loader refuses and why;
• the commands a user needs next:
  • a hosted run with --limit and a cost cap;
  • a local checkpoint with its pinned 40-hex revision;
  • the cascade's two costs;
  • plumbline report over existing artifacts.

A test loads the page's example rows through the loader, so the documented format can't drift from the real one. The page is honest that label_descriptions are loaded but not yet sent anywhere (#39).

#29: the README's hosted-vendor command sent nothing. It passed --max-cases 40 against the fixture's 105 scoreable rows, and that flag refuses rather than truncates. It now uses --limit 40, and --max-cases says in its help that it's a guard, not a truncation.

#60: the PyPI name collision. The README said pip install plumbline "will not work". It works, and installs an unrelated project.

• The README now says so and gives pip commands that install this repository.
• The local_logits missing-extra error says uv sync --extra local (and the pip form), instead of pointing at the PyPI package.
• PLAN notes the name is taken.

Checked

• The page's example rows load: 5 rows, none refused, and the score row held back from scoring.
• ruff, mypy --strict, and the full pytest suite pass.
• On the built site (12 pages), check_site_links, check_search (97 entries) and smoke_site pass, and the em-dash and -- checks are clean.

🤖 Generated with Claude Code