Bounded short-baseline image preparation

Pre-registered2026-10-30. Only six already-spent drives, the existing18 centers, and their exact ±1 capture neighbors:36 stereo pairs/72 PNG files. No new drive, LiDAR scan, fresh holdout or threshold selection. Existing timestamp files and center images remain pinned. No claim that these new frame contents were already examined; they are additional development material from spent recordings.

Stage1 reads only HEAD and ZIP central-directory metadata, verifies prior archive size/ETag and exact member presence. Payload request cap32MiB per drive and128MiB aggregate; bounds include header allowance4096 per member. Central-directory/HEAD traffic and bounded retries are separate metadata/transport overhead. Freeze exact member offsets, compressed sizes, CRC32, uncompressed sizes and request ranges in index.json BEFORE fetching any neighboring image payload. Metadata-only inventory may be read before the payload plan. No full ZIP downloads or automatic widening.

Stage2 source/index hash pins and commit required. HEAD/If-Match guard archive identity on reads; exact range lengths and ZIP CRC/size are verified. Four independent member fetch workers, fixed helper retry behavior, partial outputs/logs preserved on failure, no shrinking denominator or alternate-host retry. Retain SHA256/path for every image; workspace outputs must not overwrite prior inputs. TLS and managed proxy preserved. Check disk capacity before fetching.

Then run the SAME image-support algorithm as temporal_surface_audit on36 adjacent pairs, earlier→center and center→later, source image_02 only; retain right images for a future metric-motion study. Verify center raw grayscale pixels equal the previous audit's center images before comparing input intervals. Same ORB2000, RANSAC1px/.999, >=30 cheirality inliers and6cells screen, no retuning. Treat support as feasibility, not accurate metric motion or true overlap. Report camera timestamp intervals and every failure; no camera/LiDAR reference enters image motion.

No temporal depth mask, ICP, scale from reference extrinsics or SOTA claim. Metric camera motion remains a separate validation requirement. Any stereoVO shares stereo-depth errors and must disclose that coupling.