ds004146: MP2RAGE-only (no _T1w) — use UNIT1_denoised as the fmriprep anatomical

## What

ds004146 was skipped in the june-1 shakeout (#30): no `_T1w` files (`t1w_num=0` across all 417 subjects — **the count is correct, not an automation bug**). The anatomy is **MP2RAGE**: `UNIT1` / `UNIT1_denoised` + `inv-1/2_MP2RAGE`, plus `T2w` and `FLAIR`. BOLD is present. Valid BIDS, so this is **ours, not an upstream OpenNeuro report** (per CLAUDE.md carve-out: a pipeline that can't read a valid file is our issue).

## Want

**Figure out a way to use `UNIT1_denoised` as the T1w anatomical for fmriprep.** It's the MP2RAGE T1-weighted image; fmriprep keys on the `_T1w` suffix, so the work is finding how to present it as the anatomical. Mechanism TBD.

**Fallback:** the dataset also has `_T2w`. T2w-only anatomical is a lower-quality alternate if the UNIT1 path doesn't pan out.

## Notes

- The selector (`select-eligible-sub-ses.py`) also needs to stop skipping it. Sibling to #11 (different mechanism — that one is per-(sub,ses) aggregation).
- Raised by Austin in the 2026-06-08 meeting.