fix(baseline): read accuracy and AUROC nulls in both directions

[TMHSDigital]

Fixes #33.

The problem. Accuracy and AUROC were compared only with the upper tail of their null. A model reliably worse than the null fell far below it and was reported as INCONCLUSIVE, with the advice to collect more rows. An inverted score over 300 rows printed AUROC 0.0000 ... INCONCLUSIVE ... Collect more rows: the wrong conclusion, and the wrong advice.

What changed

• NullBand now carries a 5th percentile (p05).
• A value below it gets its own verdict, with the likely reason and the percentile it fell below:
  • accuracy: "worse than chance", because the answers are systematically wrong, usually from misaligned labels and options;
  • AUROC: "ranks incorrect above correct", because the score is inverted.
• Above the 95th percentile, or between the two, a value reads exactly as before. No existing report line changes, and the site build confirms the example report still matches its command line for line.
• METHODOLOGY now says both of these nulls are read in both directions, and why the calibration floors stay one-sided.

Checked

• New tests failed on main and pass now:
  • an inverted score (AUROC 0.0 over 300 rows) is below the null, and its statement names the 5th percentile and says "inverted", not INCONCLUSIVE;
  • all-wrong accuracy on four-way rows reads "worse than chance";
  • 25% accuracy on four-way rows is still INCONCLUSIVE.
• ruff, mypy --strict, the full pytest suite, and the site build all pass.

🤖 Generated with Claude Code