Add cau mudata-builder: the two count matrices as one MuData

[emattei]

Why

The design steps end at guide_metadata.tsv. Nothing took a sequenced
channel and produced the object an analysis works in, so that step was a
hand-written script every time.

What it does

Assembles one channel's gene and guide counts into a MuData laid out like the
IGVF CRISPR pipeline's own object, so the two can be compared field by field:
modalities gene and guide, a cell index of {barcode}_{measurement_set},
raw counts untouched, and each feature's annotation in var.

cau mudata-builder run \
    --rna counts_unfiltered/adata.h5ad \
    --guide counts_unfiltered_modified/adata.h5ad \
    --guide-metadata results/igvf/guide_metadata.tsv \
    --store gencode.v43.annotation.parquet \
    --measurement-set IGVFDS8842PXCC \
    --gtf-accession IGVFFI9561BASO \
    --outdir results/

What it deliberately does not do

No statistics, no QC, no filtering, no doublet calls, no guide assignment, no
pairs to test. Each is a judgement, and a step that makes judgements while
assembling hides them.

Three decisions worth reviewing

Cells are the union, not the intersection, with obsm recording which
modality holds each. On a real channel that is 1,022,982 barcodes against
207,274 seen by both: an intersection would discard four fifths of the object
before anyone looked at it.

Every gene type is annotated, not only protein-coding. The counts come
from a transcriptome built on the whole GTF: 20,066 protein-coding against
42,691 lncRNA, pseudogenes and small RNAs, which would otherwise be nameless.

One GENCODE, checked. A build whose store differs from the one the guide
metadata was written against is refused, so genes and guides cannot describe
different genomes. uns['reference'] then carries the GTF and genome by name,
md5 and accession, so the object answers "which GENCODE" without its run
record beside it. Only names are stored, never paths: the object is shared,
and a path carries the filesystem it was built on.

Two traps the tests pin

PAR genes appear twice -- 47 features with a _PAR_Y suffix and a chrX twin
of identical sequence -- and are carried through untouched rather than summed
or de-duplicated.

kb indexes genes as ENSG00000290825.1 while the store drops the version.
The join uses the store's own rule, which strips a trailing .N and leaves
_PAR_Y alone; splitting on the first dot instead would hand the chrY copy
the chrX gene's coordinates with nothing complaining.

Also

anndata and mudata become runtime dependencies, imported inside functions
like pysam so cau --help loads neither. Both lockfiles regenerated.

The README and two module docstrings still described location as
exon/intron/intergenic, which stopped being true in #13. Corrected here.

Verified end to end on channel IGVFDS8842PXCC: 62,757 genes and 43,766 guides,
none unannotated on either side.

🤖 Generated with Claude Code