Import predictions and prelabels as non-authoritative suggestions

[alexschmander]

Problem

Supported imports accept explicit ground truth only; external model predictions cannot be ingested as temporary suggestions with trustworthy provenance.

Current behavior

Prediction/result shapes are rejected or treated outside the supported profiles to avoid confusing non-authoritative output with ground truth.

Desired behavior

Add explicit profiles that import bounded predictions/prelabels as suggestions, never as completed annotation truth.

Acceptance criteria

• Define supported source schemas, confidence fields, task/class mapping, coverage, and model provenance.
• Validate and bound all coordinates, counts, sizes, paths, and structured input before planning.
• Keep imported suggestions non-authoritative until an annotator explicitly accepts them.
• Persist trusted source/model provenance on acceptance and preserve replay/statistics semantics.
• Cover partial/missing predictions, duplicates, thresholds, incompatible tasks, publication recovery, and UI selection.

Scope

Prediction/prelabel import profiles and suggestion delivery; not arbitrary model execution.

Implementation notes

Keep ground-truth import behavior unchanged and coordinate with exact accepted provenance in issue #40.

Tracking source

docs/tracking/feature-requests.md:42

Prediction or prelabel import as non-authoritative suggestions with model provenance.