fix(report): choose the cascade threshold on one half and score it on the other

[TMHSDigital]

Fixes #30.

The problem. The cascade chose its threshold and reported its coverage and cost on the same rows, so the printed figures were its in-sample best case. With a temperature recommended those rows were the eval split. With none, it was every row, while the text still said "held-out". On 300 synthetic rows with recalibration refused, it printed threshold 0.943 ... over 300 rows, covered accuracy 1.000.

What changed

• The threshold is now chosen on the fit half and scored on the held-out half. That is the recalibration split when there is one, or a split made the same way (make_split, same fraction and seed) when there is not.
• When a temperature was recommended, both halves are on the recalibrated scale, so the held-out rows are unseen by the temperature and by the threshold.
• The 200-row minimum counts held-out rows, so a threshold now needs at least 400 scored rows. This changes which runs print a threshold, and the cost and coverage printed with it.
• The report says where the threshold was chosen and where it was scored ("chosen on 300 rows and scored on 300 held-out rows"). PLAN says the same.

Checked

• New tests failed on main and pass now:
  • 300 calibrated rows (no temperature, 150 held out) now refuse a threshold and say why;
  • 600 rows name where the threshold was chosen and where it was scored.
• Every existing cascade test still passes, including the escalate-everything wording from fix(cascade): consider escalating everything, and price correct answers over priced rows #80.
• ruff, mypy --strict, the full pytest suite, and the site build all pass. The example report supplies no costs, so it is unchanged.

🤖 Generated with Claude Code