Benchmark TimesNet and SGN on real REDD data

[nipunbatra]

Campaign definition

• register TimesNet and SGN as distinct periodic-2D and subtask-gated model families
• advance the single immutable nilmtk-contrib pin to c130293e24e16817b9859d1b78ae18bd988b1219
• refresh the lockfile and CPU/CUDA release build contexts to the same reviewed contrib commit
• document that both models share one batched campaign image
• publish six immutable real-REDD result bundles and regenerate the living leaderboard

Real T0 campaign

• task: corrected-t1-redd / fridge / 60 s / sequence 299 / 3 epochs / limit 3960
• seeds: 10, 20, 42 per model
• hardware: NVIDIA A100-SXM4-80GB
• runner: 106af6b
• contrib: c130293e24e16817b9859d1b78ae18bd988b1219
• image: nilmbench:t0-106af6b-cuda
• digest: sha256:4b233308c342556efd0be0d88f24586d2808cc9b7f86cde1d52cafccacb4c425
• TimesNet: rank 6, MAE 53.54 +/- 3.48 W, F1 0.728 +/- 0.041
• SGN: rank 8, MAE 56.58 +/- 5.61 W, F1 0.636 +/- 0.015
• leaderboard: 63 result bundles, 21 models, one cohort, contiguous ranks 1-21

Verification

• 130 tests passed; 1 dataset-dependent test skipped
• focused ruff and uv lock checks passed
• CI cpu-smoke passed
• generated leaderboard JSON and CSV reproduce byte-for-byte
• a separate six-job A100 rerun reproduced every MAE and F1 value exactly for all model/seed pairs
• all bundles have verified source, image, hardware, dataset checksum, and result-ID provenance