Fix gblsr-scalar trained_steps -> 2000 (native matched-budget; was hardcoded 1000000)

Add arXiv + DOI badges to the model card

Remove internal source_run field from config.json (experiment-tracking jargon)

Flatten model dirs: drop redundant seed0/ level (seed recorded in config.json)

Add GB-LSR weights (native + arbitrary-scale SR variants)

initial commit