feat(site): score your own pasted predictions against their floor

[TMHSDigital]

Fixes #56.

What it does. A new "Your own predictions" section on the explainer, after the worked example. You paste rows, and it shows the measured ECE, the floor for exactly those predictions, and the verdict a report would print, word for word. The math reuses what floor.js already has: ece, calibrationFloor (seed 0, 2,000 draws, equal-width bins at the chosen bin count) and statement. That's the same path as ece_figure in the Python.

Input format.

• One row per line: a probability, then an outcome (1/0 or true/false).
• Separated by commas, semicolons, tabs, or spaces. Fields may be quoted.
• A first line that names its columns is skipped as a header. So are blank lines and one trailing separator.
• Up to 20,000 rows.

Errors. The new parsePredictions in floor.js is a pure function, exported so it can be tested. It reports each problem as { line, message }:

• a probability that is not a plain number (this includes NaN, Infinity and 0x1)
• a probability outside 0 to 1
• an outcome that is not 0 or 1
• the wrong number of columns
• no rows at all
• more than 20,000 rows

Nothing is scored until every row reads. The status line (aria-live) states the first problem, and a list shows up to ten. The box gets aria-invalid and is described by both.

Performance. Scoring runs only when you press "Score these rows", never on each keystroke, and it runs in the worker with progress and Cancel. It falls back to the main thread where a worker can't start, as the other tools do.

Privacy. The section says plainly that nothing pasted leaves the page: it isn't uploaded, stored, or put in the address. The copy button shares the result line only, never the rows.

A built-in self-check. A "Use the worked example's rows" button fills in the example's 105 rows. Scoring them reproduces the report's line exactly.

Parity.

• floor_golden.py exports four pasted texts: CSV with a header; tabs with CRLF, true/false and a blank line; spaces with the ECE exactly on a rounding tie (0.03125, printed 0.0312); and semicolons with quotes, beside a tie (0.0012499..., printed 0.0012). Each is scored by ece_figure, and --check covers them.
• check_floor_parity.mjs parses and scores each one. It requires the same row count, ECE, mean and p95 within 1e-9, and an identical verdict. It also pins 10 parser cases.
• I broke the parser on purpose (read true as false), and the check failed with 3 disagreements.

Smoke test. Two new browser checks:

• A bad row (line 3, probability 1.4) is named, the box is marked invalid, and no result is shown.
• The example's rows, pasted, give exactly the report's line, and the address doesn't change.

Checks run. ruff, format, mypy --strict and pytest pass. floor_golden.py --check passes. On a fresh build, the parity check (with the worked example), links, search, and smoke (8 of 8, no console error, no CSP violation) all pass, as do the CI prose checks. The fixture grows by about 20 KB.

🤖 Generated with Claude Code