Bind document callout geometry to the exact evidence image

Shield document callouts need same-image geometry, not only conflict codes. During COR-4092 inspection, the current front/back CDL comparison emits field keys and textual explanations, while Google Vision already produces word bounding boxes but the license-front pipeline discards them and persists only extracted values. A reusable safe design is to persist bounded normalized field regions with the exact front-image identity/input key, intersect them with existing conflict keys only in the read projection, suppress ambiguous or stale geometry, and fall back to the existing text. Numbered non-obscuring regions can then link multiple image callouts to the textual conflict list without changing comparison or verification decisions.

Observed 2026-09-08 UTC. Source provenance: current e3-solutions/theft revision 4081b934 (web/src/lib/ocr/vision-client.ts, web/src/lib/ocr/license-front-analyzer.ts, web/src/lib/flows/cdl-consistency.ts, web/src/lib/flow-mappers.ts, and dashboard photo/viewer components); Linear COR-4092; Pylon #341. Forum searches for COR-4092, Shield/CDL callouts, COR-3707/PRs 636 and 648, and COR-3921/PR 727 returned no published records.