pack_reference.sh: reject an empty reference file, not just a missing one

[mikessh]

Problem

scripts/pack_reference.sh validates that each required reference file is present in the asset (grep -q) but not that it is non-empty. A build-db run without IgBLAST emits 0-byte alleles.*/markup.*, which pass the presence check, ship in arda-reference-vdj.tar.gz, and then crash every fresh pip install at load time:

polars.exceptions.NoDataError: empty CSV
  at arda/annotate/reference.py:load_reference

v2.5.7 shipped exactly this. The published asset was re-uploaded good ~3 h after the release (verified: it is now byte-identical to the committed reference, all 5 organisms, 345 J+C scaffolds), but nothing in CI would have caught the empty asset — the guard just wasn't there.

Fix

A 1 KB size floor per required file, measured by extracting the packed member to stdout so it reflects what actually ships:

sz=$(tar -xzf "$OUT" -O "vdj/$org/$f" | wc -c)
[ "$sz" -gt 1000 ] || { echo "ERROR: vdj/$org/$f is empty ($sz B) ..."; exit 1; }

• Portable across bsdtar (macOS) and GNU tar (CI) — tested on both paths.
• Can't false-trip: smallest real file is rat/cdr3_anchors.tsv at ~24 KB, a 24× margin under the 1 KB floor.
• Runs in the release path: pack_reference.sh executes in the reference-asset CI job before gh release upload, so an empty reference now fails the release instead of shipping.

Test

• Positive: packing the real committed reference reaches OK, exit 0 (53 members, 2.8 MB).
• Negative: emptying one required file → ERROR: vdj/human/markup.aa.tsv is empty (0 B) and non-zero exit before the OK summary.

🤖 Generated with Claude Code