Correct latest-label tiebreak, harden NDJSON/CLI, honest summary wording

[wietzesuijker]

An audit pass over the export/flatten core. All behaviour-preserving except the latest-label fix (a real edge-case bug) and the CLI summary wording. Tests go 18 → 27; ruff + mypy strict clean.

Correctness

• Latest-label selection on missing timestamps. _latest_label used max(labels, key=created_at), where a label without created_at sorts as "" and always loses to a timestamped one — so a reviewer's label exported without a timestamp could be silently overridden by the annotator's. Now: if every label has a created_at, newest wins as before; if any is missing, fall back to export order (last = newest). Rare in practice (Labelbox usually populates created_at), but it's the exact failure the function exists to prevent.

Hardening

• read_export_file reports the line number on a malformed NDJSON line instead of a bare JSONDecodeError.
• The CLI validates --status against WORKFLOW_STATUSES and fails at the boundary with a clear message rather than forwarding a typo to the API.

Honest output

• The CLI printed "<n> reached unlabelled", but that count includes never-reached ToLabel rows (rows no annotator ever opened). It now prints "<n> without labels"; the field comment is corrected to match.

Tests

Lock the documented FeatureRow contract: relationship annotations, the classification/geometry "unknown" fallbacks, and the missing-API-key error — all previously unexercised.

Open questions (your call — not changed here)

• Summary.n_reached_unlabelled is misnamed (it counts all rows without labels, not just reached ones). Renaming to n_without_labels is the honest fix but a breaking API change for a published package — deferred to a major version.
• Missing-timestamp tiebreak now falls back to export order. If you'd rather treat a missing created_at differently (e.g. drop such labels, or surface a warning), say so — it's a semantic choice.