fMRIPrep: mris_expand aborts with a buffer overflow when the nipype workdir path is long (FreeSurfer)

`sub-010001 ses-02` ran 3h54m on the 2026-09-04 OpenNeuroStudies shakeout and failed. FreeSurfer's `mris_expand` aborted while building the right-hemisphere midthickness surface; `--stop-on-first-crash` ended the run there.

```
Node: _midthickness0
Cmdline: mris_expand -pial .../rh.pial -thickness -thickness_name .../rh.thickness .../rh.white 0.5 .../rh.midthickness
Stdout: reading pial surface ... using distance as a % of thickness ... reading thickness...
Stderr: *** buffer overflow detected ***: terminated
Aborted
Traceback: RuntimeError: subprocess exited with code 134.
```

## Facts

| | |
|---|---|
| Job | `63990909_1`, 8 CPUs, 40 G grant, `--mem-mb 32000` |
| `max_rss` | `14539284K` = 13.9 GiB (**of a run that died partway** — a lower bound, not a completion figure) |
| `time_elapsed_raw` | `14037` s = 3:53:57 |
| `exit_code` | `1:0` (babs's zip script; the `134` is `mris_expand` inside it) |
| `has_results` | `False` |
| Anatomicals present | `acq-mp2rage_T1w`, `inv-1_mp2rage`, `inv-2_mp2rage`, `T1map`, `acq-lowres_FLAIR`, `defacemask` |
| Stage | `--level minimal`, unchained, so it ran its own FreeSurfer recon from raw |

`mechababs iterate` routed the cell correctly: `1 job(s): 1 submitted, 0 done, 1 failed -> fail`, flagged and not merged.

## What is and isn't known

**Known:** recon-all got far enough to produce `rh.white`, `rh.pial` and `rh.thickness` — `midthickness` sits downstream of it in `anat_fit_wf/surface_recon_wf`. So the failure is not "recon-all fell over"; it is `mris_expand` aborting while reading surfaces recon-all had already written.

**A glibc fortify abort is a memory-safety bug in `mris_expand` regardless of its input.** Something about these surfaces triggers it; a well-behaved program would have errored, not overflowed a buffer.

**Not known:** whether the input surfaces are pathological. The dataset's only T1w is MP2RAGE-derived (no combined/denoised UNI), and uncombined MP2RAGE carries strong background noise that can yield poor surfaces — but that is a hypothesis fitting the file list, not something this run establishes. It also ran with `--skull-strip-t1w force`, which is worth re-examining for this anatomical.

**The evidence to settle it no longer exists.** babs's `participant_job.sh` traps `EXIT` and `rm -rf`s the per-job scratch on every exit path, so the fmriprep crash file, the recon-all log and the surfaces themselves went with the failed job. Only the SLURM `.o`/`.e` survive. Filed separately.

## Next

- [ ] Re-run one subject with the scratch preserved, then `mris_info` the surfaces — that is what separates "degenerate surface" from "`mris_expand` bug on a valid one". Blocked on the scratch-preservation issue.
- [ ] Decide the MP2RAGE anatomical handling. Shares a decision with #48 (ds004146, MP2RAGE-only, use `UNIT1_denoised`), though the failure mode differs: #48 has no `_T1w` at all, here one exists and fmriprep gets far enough to crash on it.
- [ ] If it turns out to be a FreeSurfer bug, report upstream.

Not an OpenNeuro platform problem — the content fetched fine and the files are valid; this is a tool failure on unusual-but-legal data, so it stays ours.

Found on the 2026-09-04 OpenNeuroStudies shakeout (campaign `c3`, Unity). MRIQC on the same subject/session succeeded and merged (30.2 GiB, 40.5 min).

TODO before close:
- [ ] should be filed upstream freesurfer

Start with participant_job.sh and the separate scratch-preservation issue; rerun one subject with per-job scratch retained, then inspect rh.white, rh.pial, and rh.thickness with mris_info. Done means determining whether the surface input is pathological or mris_expand is faulty, then filing upstream if the FreeSurfer bug is confirmed.

Written by the indexing model from the issue text.