feat(custom): orfnormalise + orfmerge modules + orftable_fasta_gtf_buildorfcatalogue subworkflow

[pinin4fjords]

What

Ribo-seq experiments measure which open reading frames (ORFs) on the transcriptome are actively translated by ribosomes. Five widely-used callers - RiboCode, Ribo-TISH, Ribotricer, Rp-Bp, PRICE - each emit ORF predictions in a different table format with different score semantics and classification vocabularies. There's no upstream way to combine them into one cohort catalogue.

This PR adds three components that fill that gap:

• custom/orfnormalise - parses one caller's output into a unified BED12 + sidecar TSV. The caller (one of ribocode, ribotish, ribotricer, rpbp, price) is supplied as a val input.
• custom/orfmerge - class-aware clustering of normalised calls across callers and samples into one cohort catalogue.
• orftable_fasta_gtf_buildorfcatalogue subworkflow - composes the two with bedtools/getfasta + seqkit/translate to produce a catalogue AA FASTA.

Bundled because the merger's schema is defined by the normaliser, and the subworkflow only makes sense once both modules exist.

Harmonised schema

orf_class is the union vocabulary across the five callers:

Class	Meaning
canonical_cds	Maps to an annotated CDS (or truncated / extended variant)
uORF / dORF	ORF in the 5'UTR / 3'UTR of an annotated transcript
novel_u	Novel / intergenic ORF, no annotated host CDS
smORF	aa_length <= 100, regardless of location
other	Internal / overlap / frame variants

Score direction is per-caller (p-values lower-better, Bayes factors / phase scores higher-better); the merger keeps the direction-appropriate best per cluster.

Merger clustering is class-aware:

• canonical_cds: by (transcript_id, strand) (one CDS per transcript).
• uORF / dORF / other: by (transcript_id, strand, start, end) (a transcript can host multiple distinct positional ORFs).
• novel_u / smORF: greedy reciprocal-overlap (default 0.8).

Configurable source columns + provenance

Each caller's output exposes multiple candidate columns for score (Ribo-TISH alone has FisherPvalue / RiboPvalue / TISPvalue plus Q-value variants), orf_type, and length. Defaults are in meta.yml; override per-call via ext.args:

withName: 'CUSTOM_ORFNORMALISE_RIBOTISH' {
    ext.args = '--score-field FisherQvalue'  // FDR-adjusted, not raw p-value
}

Every normalised TSV starts with a # parser_columns: ... line recording exactly which source column was read for each derived field, so downstream consumers (and reviewers) can verify provenance. Standard CSV / pandas readers skip #-prefixed lines.

Example output

Cohort catalogue TSV (called_by_* / score_* track which callers contributed):

orf_id        chrom  start     end       strand  gene_id          orf_class      aa_length  called_by_ribotish  called_by_ribocode  score_ribotish  score_ribocode
orf_00000001  20     16731702  16741005  +       ENSG00000125870  canonical_cds  224        0                   1                                   0.0101442
orf_00000002  20     17570208  17607146  +       ENSG00000125868  canonical_cds  164        0                   1                                   0.00112689

Subworkflow also emits multi-block BED12 (intron-skipping), orf_to_gene.tsv (many-to-many ORF → host lookup), catalogue AA FASTA, and a MultiQC custom-content per-class-count sidecar.

Test data

Paired with nf-core/test-datasets#2070 (five real-tool outputs sliced to <13 KB each from existing module-test outputs; full provenance in the test-data README). Tests reference the fork branch inline; will swap to params.modules_testdata_base_path once #2070 merges.

Test plan

• All 5 callers' fixtures round-trip through custom/orfnormalise with content-level assertions (non-zero aa_length, populated score, non-collapsed orf_class distribution, multi-block BED12 for multi-exon callers, provenance header present).
• custom/orfmerge chains two normaliser invocations and asserts coherent called_by_* / score_* columns.
• Subworkflow chains GUNZIP + normaliser × 2 + merger + getfasta + translate end-to-end.
• --score-field override exercised; provenance header confirms it.
• nf-core modules lint + nf-core subworkflows lint clean (only the two cosmetic Wave container-version warnings shared with custom/bed12codonpositions remain).