Fix data-split config bug and re-baseline benchmarks (split from #168)

[ZaynabEM]

Description

Split from #168 (Federated Model Training) — bug-fix-and-re-benchmark task rather than the general training/exploration work in the parent issue.

Findings so far (WG meeting, 19 Aug 2026, investigated by @gcordova10 ):

• Hard-coded data-split configuration found in the codebase
• Training/validation filtering mismatch identified
• Ablation studies show camera features currently contribute minimally to results
• Constant-velocity baseline currently outperforms the learned models — ablations point to the data/config issue above rather than an architecture problem

Findings — 25 Aug 2026 (@FLagbusted's run, reviewed by @gcordova10 ):

• @FLagbusted's run confirmed again that the constant-velocity baseline beats the learned model
• @gcordova10's review found: different input was used at each trial, so the code needs to be run consistently before results are meaningful; open questions on how the output was obtained, a right/left trajectory discrepancy, behavior across branches with different ego-motion dropout rates, and how input is supplied differently between training and validation — this lines up with, and sharpens, the train/val filtering mismatch above
• Best BEV fix so far reaches ADE 0.86
• Metric validity flagged directly: ADE compares against human maneuver, which may not be the right basis for comparison when the failure originates upstream in perception/camera rather than in the planner. Route/map/camera ablation (including camera features not fed at all) shows the metric's sensitivity to this. ADE could need to be changed for this aspect, meaning this may call for a new or modified metric
• Proposed directions: a metric that penalizes off-road predictions, ablations broken out by track position (front/right/left, on-road vs. off-road), and a more direct way to evaluate perception quality rather than relying only on end-to-end trajectory error
• Training noise noted: 3 seeds used, compared night by night
• Results expected within the week

Identified To do List:

• Fix the hard-coded data-split configuration and the train/val filtering mismatch
• Re-run the benchmark suite (same 90:10 split intent as Federated Model Training #168) with consistent inputs across trials once fixed
• Confirm whether the gap vs. the constant-velocity baseline closes
• Evaluate whether ADE/FDE need to be revised (not just re-run) to fairly capture camera/perception-driven failures, per the 25 Aug discussion