[Vincent-1512]

Problem

Product docs capped the UI at two edits and the API at three total candidates, without a multi-source selector or position-bias handling.

Changes

• Specify original plus 0-5 styles (K defaults to 4; configurable 4 or 5), guardrails, deduplication, and preview max-min diversity.
• Add learned as a planned source only; it is not implemented.
• Extend OpenAPI with learned, optional style_id/style_label, up to six candidates, and optional shown_order.
• Store shown_order in SQLite and update API design, README, architecture, and team docs.
• Keep the VLM output cap at two; do not change src, VLM prompt/schema, or DOCX files.

Verification

• PyYAML parsed OpenAPI 3.0.3: 6 paths.
• Validated 12 response/component examples against schemas.
• Applied SQLite schema in-memory; checked learned/style fields, shown_order ordering, and cascade deletion.
• git diff --check clean.

Limits

• Selector and backend/app are not implemented.
• Delta E 2000 threshold 3.0 and label thresholds are unvalidated starting points.
• learned has no data, model, or training; K=2 for silhouette/memorial is a proposal.
• keep_original.recommended remains for compatibility; UI must not use it to rank styles.

Team decisions

1. Confirm Delta E threshold and thumbnail normalization.
2. Confirm style_id taxonomy and label thresholds.
3. Approve K=2 cap and goal guardrails for silhouette/memorial.
4. Decide when to remove keep_original.recommended.
5. Define behavior if preview distance cannot be computed.