feat(query): filter by sequencing run accession; fix NA reaching download --from

[thanhleviet]

Answers a community question: "is it possible to query ATB using SRR/ERR accessions? Or only biosample IDs? I only have SRR/ERR accessions for ~6000 genomes."

Today it is sample accessions only. run.parquet already ships the mapping and is already a core table, so the data was there - nothing exposed it.

atb query --runs / --run-file

atb query --run-file runs.txt --has-assembly \
  --columns sample_accession,aws_url -o urls.tsv
atb download --from urls.tsv --parallel 8

Run accessions are resolved to sample accessions in the CLI layer, before the engine is chosen, so the SQLite index fast path and the parquet scan both see a plain sample filter. No schema change, no reindex, no second code path to keep in sync. Verified: both engines return identical rows for the same run list.

The resolution is reported rather than silently applied:

Resolved 6000 run accession(s) to 5981 sample accession(s)
5,840 result(s)

Those three numbers differ for real reasons, all covered in the guide:

• A run may carry several samples. Multiplexed runs are stored as a comma-joined list in sample_accession - 3,651 rows, hiding 10,037 distinct samples. All of them are expanded.
• Not every resolved sample is in ATB. So the result count can be lower.
• If nothing resolves, the query fails rather than returning all 3.5M rows. An unresolved filter that silently collapses to "no constraint" is the worst possible failure mode for a filter.

Runs combine with --samples as a union, and with every other filter as an AND. The pass column in run.parquet is deliberately ignored - resolving an accession is a lookup, quality filtering belongs to the query flags.

Fix: NA reaching download --from

Found while testing the above. 464,368 rows carry the literal string NA in aws_url for samples that were never assembled. parseCSVURLs treated it as a value, so atb download --from requested .../NA.fa.gz once per row - a guaranteed 404 each. Confirmed by HEAD probe that those samples have no object on S3 at all.

Now skipped and counted:

Skipping 1065 row(s) with no assembly (aws_url is NA)
Downloading 4775 file(s) to /home/ubuntu/atb/genomes

Verification

Built from source and run against the live database:

Check	Result
6,000 runs → samples → URLs	5,981 samples → 5,840 in ATB → 4,775 downloadable, 4.6 s
Index path vs parquet path	identical rows
Multiplexed run ERR016606	1 run → 12 samples
--has-assembly vs skip count	5,840 − 4,775 = 1,065, matches exactly
Spot-checked assembly URL	HTTP 200
All-bogus run list	exits 1, does not dump the database

Tests are TDD throughout: 22 new tests across internal/runs, internal/query and internal/cli, each watched fail first. Full suite green (21 packages), go vet clean, mkdocs build --strict passes, make docs regenerated.