Wildbook only requests heatmask images. The matches (SIFT feature
overlay) and clean render types were never displayed. Removing them
cuts per-job rendering from 3 images to 1 per match annotation,
reducing a 200-job bulk import from ~100 minutes to ~33 minutes.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Instead of synchronously rendering heatmask images for every match
before returning the job result, defer rendering to the /match/thumb/
endpoint. When a user clicks "inspect" in Wildbook, the endpoint
renders the side-by-side chip image on first request and caches it
for subsequent requests.

This decouples image rendering from the identification pipeline,
allowing the callback to Wildbook to fire immediately after inference
completes instead of blocking for 30-90s of image I/O.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

The zero-score filter could empty the candidate list on queries with
no positive matches, causing downstream assertion failures. Fall back
to the single best-ranked result so Wildbook always has at least one
candidate to display.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

HotSpotter's yellow heatmask overlay requires the ChipMatch spatial
data which is only available during query_chips_graph. Render
synchronously for HotSpotter (proot=vsmany) to preserve the overlay.
MiewID, PIE, and other algorithms continue to use on-demand rendering.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>