[NingyuSUN]

Adds the first expert-review round for #23: one workbook that experts fill in, plus the importer that turns their answers into each review's own format.

The workbook

Each part matches a field of expertise, and an expert does only the parts in their field.

Part	Items	Feeds
A · Variants	40 ClinVar germline classifications, Sept 2023: 20 divergent and 20 control, not marked as such	ClinVar review kit (#23), scoring of the model reviewers (#22)
B · Paper claims	the 60-claim extraction expert sample	natural extraction errors with taxonomy codes (#21)
C · Single-cell	a fresh audit sample: 59 auto-admitted, 10 others, not marked as such	audit of what bioevidence admits (#24)

• Every answer is a dropdown with plain options: "yes / no / unsure", and "usable as is / needs a curator's check / not usable".
• The options live on a hidden sheet, so options containing commas don't break the dropdown.
• The start sheet gives six rules in English and Chinese.
• Blinding: the workbook shows no model, route, validator decision or reference label. Parts B and C get new item codes. The seed, the item key and the audit manifest stay in the maintainer's private folder.

The importer

import_round.py writes:

• part A as protocol annotations, through import_sheets.convert;
• part B as rows in the extraction sample's labelling format;
• part C as protocol annotations that bioevidence review audit-score scores against the private manifest.

It rejects half-answered rows, unknown options, a missing reviewer ID and a repeat import of the same reviewer. Nothing is written unless every part is clean.

Checks

• A round trip on the real round passed: built, filled as an expert would, imported, then checked with review check and scored with audit-score. 120 ClinVar annotations, 60 claim labels and 69 audit annotations came back.
• 3 new tests cover blinding, dropdowns, the 20/20 balance, the import, partial and invalid answers, and repeat imports.
• ruff, mypy and pytest pass, with 98.6% coverage. mkdocs build --strict passes.