Resolve companion/auxiliary files (checksums, indexes) — propagate from parent or tag by type, genomic dims not_applicable

[NoopDog]

~64K of the ~80K currently-unclassified files are non-primary-data, and a large chunk are companions of a real file — .md5 checksums (14,179), plus indexes (.bai/.fai/.gzi). Today they're silent gaps. They should be explicitly resolved, like classify_index_files.py already does for .tbi/.crai.

Idea (from the field, 2026-07-31)

"Classify the checksums as checksums and then call everything else NA, or link it to the thing it is a checksum of, like we do with index."

Two acceptable resolutions per companion file:

1. Tag by type — .md5 → data_type checksum (or similar), genomic dimensions (reference/modality/assay/platform) not_applicable. A positive determination, not an unknown (cf. Distinguish 'unreadable' (fetch failed) from 'read but no expected content' — VCF/FASTQ mislabel a clean read as content_unreadable #299 unreadable-vs-not-this-type).
2. Propagate from parent — x.vcf.gz.md5 inherits x.vcf.gz's classification (as .tbi/.crai already do). Better when the parent is classified; keeps the companion consistent with what it accompanies.

Likely: propagate when the parent is present/classified, else tag by type.

Scope

• Checksums: .md5 (~14K) — strip the trailing .md5 to find the parent; propagate, else tag checksum.
• Indexes: extend the existing .tbi/.crai propagation to .bai/.fai/.gzi (~2.5K). (This supersedes the index-propagation item in Add classification rules for the unhandled data-format tail (~16K files: bigWig, GFF3, indexes, chains, h5ad, PAF…) #311; leave Add classification rules for the unhandled data-format tail (~16K files: bigWig, GFF3, indexes, chains, h5ad, PAF…) #311 for actual data formats — bigWig/GFF3/etc.)
• Standalone docs (.txt/.log/.pdf/README, ~46K) are NOT companions — out of scope here; they stay honestly unclassified (or a later "documentation" tag with genomic dims not_applicable).

Why it matters

• Turns silent gaps into positive not_applicable/propagated answers → the residual "unclassified" count then means "real gap," not "auxiliary file."
• Consistent with the existing index-propagation pattern and the accuracy-over-coverage principle (a checksum having no reference is a fact, not a failure).

DoD

• .md5 resolved (propagate-from-parent, else checksum + genomic dims not_applicable).
• .bai/.fai/.gzi propagate from parent like .tbi/.crai.
• Companion resolution reuses/extends classify_index_files.py rather than a new one-off.
• Coverage re-measured; unclassified drops toward genuine-gap-only.

Relates to #311 (data-format tail), #299 (resolved vs unknown), #301 (producing method).

[NoopDog]

Cross-link from the #314 self-consistency linter (PR #319, merged):

The auxiliary_inert rule (a checksum/log file must have modality/reference/assay/platform all not_applicable) deliberately excludes index, because index files intentionally inherit their parent's classification via propagation — so flagging them would flag ~210K intentional records, not bugs.

Whether index files should be inert or inherit is exactly the question this issue (#312) resolves. Once decided:

• if index files should be inert → extend auxiliary_inert (or add an index_inert rule) in src/meta_disco/rules/consistency_rules.yaml to cover data_type: index;
• if they should inherit → leave the exclusion, and document it here as the intended behavior.

No action needed in the consistency linter until #312 lands.