Improve duplication risk scoring in the fit score

[yoldaolmak]

Problem

Images with similar visual characteristics shouldn't be placed next to each other in the same article. The current fit score doesn't include a structural similarity check between selected candidates.

Scope

• src/pictova/engine/quality.py and the fit score math documented in docs/concepts/fit-score.md.
• Add a basic structural similarity signal (e.g. perceptual hash distance or color-histogram comparison) between candidates competing for adjacent slots.
• Keep it dependency-light: Pillow + numpy are already available.
• Unit tests with fixture images that are near-duplicates vs. clearly distinct.

Acceptance criteria

• Near-duplicate candidates are penalized when targeting adjacent placements.
• The signal is a documented, weighted component of the fit score, not a hard filter.

Difficulty: High · Credentials: None.

[yoldaolmak]

Maintainer update: this is a design-sensitive core change, not a first contribution task. Before implementation, propose the candidate comparison signal, its weight, and fixture strategy in a comment. The score must remain deterministic and explainable in the resulting Fit Score reason.