[iskandr]

Summary

Adds on-demand viral reference downloads and gives ORFdata one consistent API
for collections, species, sources and protein inputs without exposing storage
machinery as domain concepts.

• ViralCollection describes one published collection. The same object drives
  discovery, inventory reads, source verification, downloads and offline opens;
  adding a species or source is registry data rather than another code path.
• collection.inventory() and collection.source_summary(source) return pandas
  frames with declared columns. The primary source is the default everywhere.
• download_viral_references(), open_viral_references() and
  select_viral_references() work for every collection and return the same
  ViralReferenceFiles interface.
• list_data() -> tuple[DataStatus, ...] discovers bundled, downloadable and
  buildable data without implying that every item is a download.
• A Proteome contains ProteinSequence values in .proteins.
  ProteinSet is the general sequence-only input for a selected component;
  with_protein_set() records donor/transgene use and with_viral() adds a
  sourced host association. ncORFs with coordinates and evidence remain
  ORFCatalog records and produce the same ProteinSequence output.
• Checked-TSV declarations, source checks, fetchers and cached source documents
  are internal implementation details rather than package-root vocabulary.

Downstream producer contract

The open Varcode, Isovar and Vaxrank requests are covered directly:

• exporter/wrapper names are provenance, not source identity; direct and
  wrapped copies merge once and retain all paths in exporters
• an external protein may retain taxon_id=None; admitting it to a known host
  requires compatibility="allow_unknown" and does not rewrite the source fact
• producer version, provider, export paths, exact reference, source checksum,
  ORF category and true/false/unknown boundaries appear in safely escaped FASTA
• exact-sequence FASTA deduplication uses a neutral group header and maps back
  to every contributing source protein
• a selected donor/transgene set retains its own taxon/reference/checksum and a
  separate sourced assertion for the target proteome

Proteome snapshots use schema v5. Loading v1-v4 remains supported; v4 producer
IDs and graph links are normalized away from exporter-specific identity while
retaining the exporter path.

Viral reference data

orfdata download viral --species human installs gdc-viruses, the 200-record
GDC collection, in identifier-compatible genome, transcriptome and proteome
layers. viral_genome() opens the collection or a selected subset as a
PyEnsembl genome for Varcode, Isovar, Topiary and Vaxrank.

Every fetched source record must reproduce the pinned source identifier,
sequence length, MD5, SHA-256 and coding-feature count. GDC's HPV16 contig is
therefore taken from PaVE's HPV16REF, whose sequence reproduces the published
GDC MD5, rather than the different GenBank accession printed beside it. A test
registers a mouse collection and downloads it end to end without a new loader.

Named YAML subsets keep panel membership, host association and tumour mechanism
separate. Optional alternate annotations apply only to an identical molecule.
transcripts.tsv records initiation and recoding, and
viral_annotation_survey() reports annotation completeness without a built-in
gene list.

Validation

• ./test.sh: 826 passed
• ./lint.sh: passed
• ./docs.sh: passed
• python -m build and python -m twine check: passed
• focused producer identity, unknown-taxon, generic ProteinSet, legacy v4
  migration and FASTA safety tests: passed