World-model quality benchmark (JEPA + Reactive vs Combined)

[ShauryaVM]

The existing speed benchmark covers FPS/latency for Combined, but we didn’t have a quality side for the world-model path. This adds that.

What’s in here:

• JEPA reconstruction metrics (per-horizon L1/L2/cosine), plus a null-predictor baseline and relative improvement
• A Reactive vs Combined trajectory-impact helper (same general shape as the faithfulness helpers)
• Open-loop ADE/FDE pairing when GT controls are available
• CLI under Model/evaluation/world_model_quality_benchmark.py (--synthetic / --impact)
• Short notes in QUALITY_BENCHMARKS.md and a link from the speed-benchmark README

Action-conditioned WM wiring and reasoning-confidence planner coupling are intentionally left for follow-ups.

How I tested

• pytest Model/tests/test_world_model_quality.py
• python Model/evaluation/world_model_quality_benchmark.py --synthetic
• ruff on the touched files

[riita10069]

Under the contribution standard proposed in #198, I consider this PR AI slop in its current form. It presents itself as a quality benchmark, but only evaluates synthetic target-plus-noise tensors and a randomly initialized model; it reports no result from real data or a trained checkpoint. This is not ready for review; please provide the required real-world validation before resubmitting.

[ShauryaVM]

Pushed a follow-up so this is no longer a random-init / synthetic-only smoke.

What changed

• Default path is now --trained: Combined (IL + JEPA) is trained, then we report JEPA recon and Reactive vs Combined ADE/FDE on the same trajectory target.
• --shard-dir / --checkpoint is there for a packed KITScenes/L2D partition when you have one (--train-steps 0 to eval a checkpoint as-is).

Numbers from this machine (CPU, mock backbone, 12 steps, seed 0; JSON in Model/evaluation/results/world_model_quality_trained.json):

	Reactive	Combined	Δ (C−R)
ADE@3s	4.160	3.971	−0.189
FDE@3s	11.000	10.428	−0.572

Combined loss 0.462 → 0.389. JEPA relative improvement vs a zero predictor is 0 on random frames (expected without real video). I am not claiming a KITScenes paper number — this is a trained Combined A/B so the ADE pair is defined. Happy to re-run on a shard you point me at.