test: add regression test for RFDETRSegNano PE interpolation at custom resolution

[Copilot]

RFDETRSegNano(resolution=1008, num_classes=2) raised RuntimeError: size mismatch for backbone.0.encoder.encoder.embeddings.position_embeddings when loading pretrain weights — the checkpoint's 26×26 PE grid (default 312px resolution) wasn't being interpolated to match the model's 84×84 grid (1008px / patch_size=12).

The fix was already present in the current codebase: inference._build_model_context routes through load_pretrain_weights → interpolate_position_embeddings, and ModelConfig._sync_pe_with_resolution correctly propagates positional_encoding_size = resolution // patch_size when a custom resolution is supplied. However, there was no end-to-end test covering the segmentation model path.

Changes

• tests/training/test_load_pretrain_weights.py — adds test_rfdetr_seg_nano_loads_default_pe_checkpoint_at_custom_resolution to TestL1FacadePEInterpolationEndToEnd, mirroring the existing RFDETRNano test for the segmentation variant:
  1. Builds a default-resolution RFDETRSegNano (no pretrain weights, 90 classes) and saves its state dict as a checkpoint with args.segmentation_head=True, args.patch_size=12
  2. Reloads at resolution=624 (num_classes=2), triggering PE interpolation from 26×26 to 52×52
  3. Asserts the backbone position_embeddings tensor shape is [1, 2705, 384] (52²+1 tokens)

# Before fix (v1.6.5.post0): _load_pretrain_weights_into in detr.py skipped interpolation
# RuntimeError: size mismatch for backbone.0.encoder.encoder.embeddings.position_embeddings:
#   copying a param with shape torch.Size([1, 677, 384]) from checkpoint,
#   the shape in current model is torch.Size([1, 7057, 384]).

# Now works:
model = RFDETRSegNano(resolution=1008, num_classes=2)
model.train(dataset_dir=..., epochs=10, batch_size=1)