Plant trees from aerial imagery and ship both vision detectors

[vanlueckn]

Summary

• Tree placement from imagery: a model class marked kind: Tree in ai.ron plants finds as ordinary rows in the line's tree list (via TreeSource) instead of objects bolted to the track graph — selectable, deletable, and undoable exactly like hand-planted trees (vision::canopy, route-editor/src/ai.rs).
• Measured crowns drive size: the species is drawn from those whose crown is within 1.5× of the measured span, then scaled (clamped to 0.6–1.5) so the tree matches the crown; all 84 mods/trees objects gained a footprint measured off the built crown_LOD0 mesh by the updated tools/trees/import_mantissa.py.
• Conifer/broadleaf guess: where the model only knows "Tree", vision::canopy classifies crowns via two ratio-based features (brightness spread and red/blue warmth), thresholds calibrated on real provider imagery; both must agree.
• Species override in the dialog: "As detected" or a named stand (nadelwald, laubwald, …) which plants like the forest brush; also available headless via --stand.
• Second output layout Head::Retina: torchvision RetinaNet emits logits/offsets per anchor with no boxes; the anchor grid is rebuilt in Rust and validated against torchvision's numbers (including banker's-rounding edge cases).
• Both detectors now ship in models/ (Git LFS): yolov8n-obb.onnx (12.7 MB, AGPL-3.0, Ultralytics/DOTA) and deepforest-tree.onnx (129 MB, MIT, DeepForest/NEON). Licence implications documented in models/LICENSES.md, THIRD_PARTY_LICENSES.md, and the README; deleting the AGPL file degrades gracefully.
• Model registry changes: span field for classes where the factor-of-two size rule is wrong (crowns range 3–25 m); /models/ removed from .gitignore, LFS tracking added in .gitattributes; tools/vision/export_models.py reproduces both conversions.
• Duplicate suppression reworked: occupancy test is now a grid (ai::Occupied) instead of a list, so tens of thousands of trees don't freeze the editor; tree de-duplication uses ⅓-crown spacing (configurable "Keep trees apart").
• i18n: new keys (ai-place-trees, ai-species, ai-tree-spacing, …) added to both en and de locales.
• Docs updated: README (trees section, shipping table), MODS.md, PLAN.md, STATUS.md, tools/vision/README.md.

Testing

• cargo test --workspace: 1136 existing tests plus new crates/vision/tests/shipped.rs (136 lines) covering the shipped detectors — including Retina anchor-grid validation against torchvision reference numbers.
• Retina decode verified against DeepForest's own test image: Python reference finds 55 crowns, this pipeline 50 at confidence 0.3, shared boxes agree to ~1 px (~0.1 m).
• cargo test -p i18n confirms en/de key parity for the new strings.
• Canopy threshold calibration measured on provider imagery (contrast 0.10 broadleaf vs 0.22/0.13 conifer); noted in module docs as resolution-dependent (needs ≥3 m crowns to be meaningful).
• Headless run (--detect-run --stand nadelwald) writing trees into the route file: not verified in this summary — run locally if needed.
• Second-run idempotency over the same wood (grid occupancy + tree spacing): not run here.
• LFS pull on an existing clone (git lfs pull) and editor behaviour with yolov8n-obb.onnx deleted (graceful "not installed" report): not run.
• Editor UI flow (dialog, progress, undo, select/delete of planted trees): manual check recommended on a real route; not run.