fix(loaders): refuse what cannot be scored honestly, in both loaders and the pricing file

[TMHSDigital]

Fixes #44. Fixes #45.

#44: load_jevbench skipped the duplicate-id check that load_jsonl makes, so ["dup", "dup"] loaded with no refusal. Both loaders now share one check and one message.

#45, dataset loader:

• Byte order mark: a UTF-8 BOM (which Windows editors and spreadsheet exports write) made line 1 unparseable. The first line is now read as utf-8-sig.
• Not UTF-8: a non-UTF-8 file failed with a bare codec error. Lines are now decoded one by one, and the error names the file, the line and the byte.
• Labels: str() turned a label of null, true or 1 into "None", "True" or "1" and loaded an option nobody wrote, and empty or blank labels were accepted. Every label must now be a non-empty string. The public fixture's labels are all strings, and it still loads 111 of 111.
• Descriptions: label_descriptions could describe options that don't exist, and null became "None". Keys must now be options, and values non-empty strings.

#45, pricing loader:

• NaN, infinite and negative prices were accepted. They are now refused.
• as_of accepted 20260901 and 2026-W36-1 (fromisoformat takes both) and dates in the future. A future date would keep the report from ever calling a price stale. All three are now refused.
• The pricing file is read as utf-8-sig too.

The "Your own data" page lists the new refusals.

Checked

• New tests failed on main and pass now:
  • a duplicate JevBench id is refused;
  • a BOM file loads;
  • a latin-1 line is named ("latin1.jsonl ... line 2");
  • five kinds of bad label are each refused;
  • three kinds of bad description are each refused;
  • six bad pricing values are each refused;
  • a BOM pricing file loads.
• ruff, mypy --strict, the full pytest suite, and the site build all pass. The example report is unchanged.

🤖 Generated with Claude Code