ds004078: ~60 BOLD runs in one subject-level job exceeds walltime (no sessions to split on)

**TL;DR:** ds004078 has no sessions, so sub-01's ~60+ BOLD runs all process in a *single* subject-level job — too much work to finish in a reasonable SLURM walltime. It ran the full **24 h** and was CANCELLED at `TIME LIMIT`. ~60 runs/job is abnormal; this dataset needs special handling. (Hopefully 60-runs-per-job is not common across OpenNeuro.)

## Evidence
- Subject-level (no `ses-*`): `sub-01/func/` holds `task-RDR_run-1 … run-60` plus `task-rest` runs.
- minimal fmriprep started `2026-06-03 01:34`, SLURM CANCELLED `2026-06-04 01:27` (~24 h, `DUE TO TIME LIMIT`).
- Aggravated by `--use-syn-sdc` (no fieldmaps → ~81-min SyN per run; only 15 of ~64 finished in 24 h) — but ~60 runs/job will strain any reasonable walltime even without it.

## Direction (open)
Hard-dataset handling: how to process a session-less subject with a very high run count — more walltime (likely exceeds cluster caps), split the runs, or special-case it. Relates to #19 (awkward datasets) and #3 (resource estimation).

## Done when
ds004078 sub-01 processes to completion, or we have a documented approach for high-run-count session-less datasets.