Fix rotted PDB download (#85) and add a MaSIF-site CI smoke test

[DoctorDean]

What this fixes

source/data_preparation/00-pdb_download.py downloads structures via Biopython's
PDBList.retrieve_pdb_file(...), which relies on legacy wwPDB download paths that RCSB
has since retired. It now fails silently — no file is written — and the whole
data_prepare_one.sh pipeline then dies downstream with a confusing "structure doesn't
exist" / IndexError. This is the root cause behind #85 (and the symptoms in #83).

The fix is a two-line change: fetch the structure directly from the current, stable
RCSB endpoint (https://files.rcsb.org/download/<ID>.pdb) instead of going through
Biopython's retired path. It's Python 2/3 compatible and needs no new dependencies.

Verification

Ran the canonical example end-to-end on the published pablogainza/masif:latest image
with only this change (no --file workaround):

./data_prepare_one.sh 4ZQK_A     # built-in download — now works again
./predict_site.sh 4ZQK_A
./color_site.sh 4ZQK_A
# -> ROC AUC score for protein 4ZQK_A : 0.9137

That matches the value in the paper, so the whole interaction-site path is healthy again.

Also included: a CI smoke test

.github/workflows/masif-site-smoke.yml runs exactly the above on every push/PR (inside
the published image) and asserts the 4ZQK_A ROC-AUC stays ≥ 0.8. The repo currently has
no CI, which is why the download breakage went unnoticed for years — this gives MaSIF a
reproducibility heartbeat so it can't silently rot again.

Changes

• source/data_preparation/00-pdb_download.py — direct RCSB fetch (the fix)
• .github/workflows/masif-site-smoke.yml — new CI smoke test

---

Full disclosure: the diagnosis and fix were produced by Lazarus
(https://github.com/DoctorDean/lazarus), an autonomous agent that resurrects dead research
code, and then verified by hand end-to-end (the ROC-AUC above is a real run). Happy to
adjust anything to fit the project's conventions.