fix(cascade): consider escalating everything, and price correct answers over priced rows

[TMHSDigital]

Fixes #34. Fixes #35.

#34: the cascade never considered escalating every case.

• The candidate thresholds were 0 plus every observed score, and a row is covered when its score reaches the threshold. So the highest score still covered the rows that reached it, and "escalate everything" was never a candidate.
• With every case wrong, an error at $100 and an escalation at $1, it chose threshold 0.9 at $201 where escalating all three cost $3.
• ESCALATE_ALL (infinity) is now the last candidate. When it wins, the report says in words that escalating every case is cheapest, instead of printing a threshold of inf.

#35: per_correct_usd understated the cost of a correct answer.

• It divided the total over priced rows by correct answers across all rows: summarize([1.0, None], [True, True]) gave 0.5 where 1.0 is right.
• It now counts correct answers among priced rows only.

Checked

• New tests failed on main and pass now:
  • the all-wrong case now escalates all three at $3;
  • per_correct_usd is 1.0 in the case above;
  • the report says "Escalating every case is cheapest" and prints no inf. This test forces the optimum to the new candidate, because the mock refuses an accuracy below chance.
• The test that pinned the exact list of candidate thresholds now includes ESCALATE_ALL.
• ruff, mypy --strict, the full pytest suite, and the site build all pass.

🤖 Generated with Claude Code