feat(megadetector): add MegaDetector v6 as a standalone model choice

[Chouffe]

Summary

Adds MegaDetector v6 (MDV6-yolov10-e-1280) as a fourth model in Biowatch's AI Models tab. MD is a fast worldwide detector for animals / people / vehicles — no species classification — useful as a blank-filter before manual species annotation. The three MD labels flow through Biowatch's species-centric data model as pseudo-species (animal, homo sapiens for person, vehicle).

Architecture

• New Python LitServe server run_megadetector_server.py modeled on run_deepfaune_server.py minus the classifier. Loads weights via ultralytics.YOLO.
• New tarball hosted at huggingface.co/earthtoolsmaker/megadetector/main/6.0.tar.gz (117.8 MB), built by scripts/build-megadetector-tarball.py from the Zenodo-hosted upstream weights.
• New detectionOnly: true flag in mlmodels.js that suppresses the species panel in ModelCard.jsx and the model-picker dropdown.
• New megadetector branch added to parseScientificName, transformBboxToCamtrapDP, and detectModelType. Common-names resolver gets a megadetector source so animal/person/vehicle resolve cleanly; the person label is mapped to the binomial homo sapiens at the server so species tooltips and IUCN lookups fire.
• Confidence threshold: 0.2 (MD's official default).

Bonus fixes (caught during manual smoke)

• Corrupt-video false positive: get_video_metadata now sanity-checks fps/frame_count/duration so the entire-file error pipeline catches corrupt files instead of letting one garbage frame surface as "animal".
• Per-bbox confidence in observations: for MD, each row's classificationProbability now mirrors its own detectionConfidence instead of the top-level prediction score being copied across every row.
• Video winner selection for MD: new weightedVote mode in selectVideoClassificationWinner uses mean-confidence-primary voting (frame-count tiebreaker) for MD videos, so a short burst of high-confidence "person" detections isn't drowned out by a longer stream of weak "animal" detections. Classifier behavior (SpeciesNet/DeepFaune/Manas) is unchanged.

Design + plan docs

• Spec: docs/specs/2026-05-16-megadetector-model-choice-design.md
• Plan: docs/plans/2026-05-16-megadetector-model-choice-plan.md

Test plan

• JS: `npm test` (1362/1362 pass, includes 9 new MD-specific tests across resolver, mlmodels, bbox, prediction, and video classification)
• Python: `cd python-environments/common && make lint && make format && uv run pytest tests/test_megadetector_server.py tests/test_utils.py -v` — 8 e2e tests against real weights + 13 utils tests (incl. 6 new metadata sanity tests) all pass
• Manual smoke: download MD from AI Models tab → import mixed image folder + videos → confirm: card shows "Detection only · 3 categories", boxes show per-detection conf, blanks land as blanks, person frames land as "Human (Homo sapiens)", corrupt videos are skipped, DSCF0028.MP4 picks person not animal
• Regression check: import a folder under an existing SpeciesNet/DeepFaune study, confirm classifier model behavior unchanged

Notes

• ~118 MB download, reuses the existing `common` Python env (no extra env install for users who already have SpeciesNet/DeepFaune/Manas)
• License: MDv6 weights are AGPL-3.0; the upstream LICENSE is bundled in the tarball so attribution travels with the binary
• The `logo` field in `mlmodels.js` is unused throughout the codebase, so no MD logo asset is required

[netlify]

✅ Deploy Preview for lucent-yeot-0cb408 canceled.

Name	Link
🔨 Latest commit	5ce050e
🔍 Latest deploy log	https://app.netlify.com/projects/lucent-yeot-0cb408/deploys/6a0ad9497763410009f258d4