Add a stable Image alias for Dataset Viewer delivery
The same-row diagnostic shows identical embedded JPEG cells returning an unexpired HTTP 403 under Physical reference photograph while image returns HTTP 200 and fully decodes. The timestamped public observation is pinned at https://huggingface.co/datasets/maxwellinked/time-lapse-artifacts-viewer-test/blob/8ee47ec5c8fe90007aab1c1b7d6377e3d152bc3b/observations/2026-10-20-column-delivery.json . This establishes column-dependent delivery, without establishing a specific CDN/S3 implementation cause.
Add image in the second position of both existing embedded-photo Parquets. Keep Physical reference photograph, every original field/type/cell and all video references; move the legacy Image field to the final position and preserve the positions of all other original columns. Each column contains the same 971 selected Image cells and 362 true nulls. The CSV, source photo files, recording/work identities, relationship evidence, all five configurations and VideoFolder loader are preserved. Only the paired feature declaration gains the alias and new order.
Add a small offline PyArrow migration helper and seven focused contract tests so future Parquet builds preserve the workaround. The helper does no media decoding/network access and rejects in-place rewrites and existing-output overwrite. Both Parquets retain the original Image cells byte for byte. The two serialized files increase by 185,623,758 bytes in total; actual Hub deduplicated storage is not measured. This compatibility cost preserves the existing named Image API.
Independent review approved the exact six-file inventory, original-field/cell/schema/metadata invariants, null controls, retained source CSV and configurations, documentation, and all seven helper tests. Hosted row readiness and generated-image delivery remain separate live checks after publication. GitHub website inputs are unchanged by this Viewer-only correction.