Quality gate + mandatory fields

[monneyboi]

export-csv currently emits every extracted observation, with no filtering. We need to decide:

• Which fields are mandatory for a row to be exported? Today the non-blank core in practice is dataset, source_url, snapshot_id, organisation_name, person_name, position_name (with its fallback). Every other field is blank when the page doesn't state it. Should we filter incomplete rows?
• Should zavod consume every extraction, or only observations accepted by a Kolkhoz quality gate? A gate would likely be based on provenance signals rather than scoring — e.g. whether the value was extracted from plaintext vs. an image, whether it came with a page locator, whether the extracted string appears verbatim in the source document, etc.

Carries over from #4.