Single source figure exported as multiple separate image files instead of one

On a 36-page test document (mixed text/code/screenshot SAP training PDF), one figure ("Figure 102" in the source numbering) that is a single image in the source PDF is exported by Marker as multiple separate image files, each covering a portion of the original figure. The markdown output then references multiple fragment images in sequence where a reader expects one unified figure — visually broken when rendered (misaligned edges, no indication the fragments belong together).

Repro attempted, not yet reproduced: tried 3 synthetic constructions to isolate the trigger — (1) one diagram split into two image page-objects, stacked touching edge-to-edge; (2) a single image object with a large blank visual gap between two content clusters; (3) variant (1) with a stray text object sitting at the seam between the two tiles. All three were correctly merged by Marker into one exported Figure/Picture + FigureGroup/PictureGroup — none reproduced the fragmentation seen on the real document (confirmed via marker-pdf==1.10.2's own *_meta.json block_counts for each attempt). The real-document trigger remains unidentified; filing this as an observation report rather than blocking on finding the repro. Candidate untried angles: overlapping-but-not-touching tiles, tiles of different pixel dimensions/DPI, or a multi-column page layout interacting with the figure's bounding box.

Ask: when multiple image-type page objects are spatially adjacent and clearly part of one visual unit (touching or near-touching bounding boxes), consider merging them into a single exported image, or at minimum grouping them in the output so a downstream renderer/reader can tell they're one figure. Happy to help narrow down the trigger further if there's a known code path this might point to.

Environment: marker-pdf==1.10.2.

Related: #1091 (this repo's other report from the same test document, on silent image loss).