Add .tbl NCBI feature table output for GenBank submission

[gbouras13]

Closes #137.

Adds <prefix>.tbl, an NCBI feature table, alongside the existing outputs of phold run and phold compare, so a Phold-reannotated genome can go to table2asn without a hand conversion. Modelled on the .tbl Pharokka emits (pharokka.post_processing.create_tbl) so the two tools stay interchangeable in a submission pipeline.

Not produced for proteins-predict / proteins-compare, which have no contig coordinates.

What's in it

• All CDS, plus the tRNA / tmRNA / CRISPR features carried over from the input GenBank. Phold doesn't predict those, but a feature table that omitted them would describe fewer features than the .gbk written beside it.
• Incomplete CDS marked with < / >, with codon_start where the reading frame requires it. partial is plumbed through from both input routes — newly recorded from gene.partial_begin/partial_end on the pyrodigal path, and read from Pharokka's existing /partial qualifier on the GenBank path. The column is dropped at the merged_df join, so no existing TSV gains a column.
• Contig lengths passed through so a minus-strand 5'-partial CDS can be placed against the contig edge.

Validated against table2asn

Checked with table2asn 1.28.1179 (bioconda, osx-64). Three problems showed up, all of which would have hit anyone actually submitting:

/inference was rejected on every single CDS. It was being written as the raw annotation_method (foldseek, pharokka, none). INSDC requires a category from a controlled vocabulary, so table2asn returned Qualifier had bad value — 246 hard errors on SAOMS1, 15/15 on a pyrodigal run. Pharokka has the same problem, since it writes its GFF Method column here.

Category picked by measurement rather than guesswork:

/inference value	result
similar to AA sequence	InvalidInferenceValue warning
similar to AA sequence:PHROG:1215	warning — "unrecognized database"
protein motif:PHROG:1215	clean
alignment:Foldseek:1215	clean

The similar to ... sequence categories are checked against NCBI's recognised-database list, which PHROG isn't on. protein motif isn't checked, and is also the honest description: a PHROG is a protein group / profile, and Phold transfers annotation from the group its structure matched. This is the one judgement call worth your review — it's a single constant (_INFERENCE_CATEGORY) if you'd rather claim something else. CDS with no PHROG to cite get no /inference at all, rather than a bare category (itself flagged) or an invented accession.

/anticodon was rejected for Pharokka < v1.8.2 input, which writes a bare 3-letter code where NCBI wants a location ((pos:115194..115196,aa:Met)). write_genbank already warns about that input; the qualifier is now dropped with a warning instead of written knowing it's invalid. The valid location form passes through untouched.

3'-partial CDS warned PartialProblemNotSpliceConsensus3Prime. Gene callers stop at the last whole codon, leaving 1–2 trailing bases; these are now extended to the contig edge, mirroring the 5' codon_start handling.

After the fixes the only findings left are the submission metadata table2asn always wants (NoPubFound, NoTaxonID, MissingPubRequirement — supplied by the submitter's template.sbt and organism), plus BadCDScomponentOverlapTRNA on SAOMS1, which is genuine in that input: CDS_0193 (115129..115329) spans the tRNA at 115194..115265, and Pharokka's .tbl would report it identically. Zero feature or qualifier findings attributable to the writer.

Two bugs caught mid-review

Both were in my own first pass, and both surfaced only once I drove the real write_genbank instead of a hand-built dataframe:

1. Strand reaches the writer as "+"/"-" strings — write_genbank converts before returning per_cds_df. int(strand) == -1 would have raised ValueError on every real run, at the final write step.
2. per_cds_df is already in transcription order for CDS (start = location.end for minus strand, so start > end). Re-orienting that pair wrote every reverse-strand CDS backwards — no crash, just a silently wrong submission file, 78 of 246 CDS on SAOMS1.

Departures from Pharokka

• A CDS partial at both ends ("11") gets both markers, instead of falling through Pharokka's if/elif chain unmarked.
• An out-of-range codon_start warns instead of aborting — this runs at the very end of a long pipeline.

Testing

57 unit tests in tests/unit/test_tbl.py pinning the exact line shape: tab counts, coordinate order on both strands, BioPython half-open → 1-based conversion, all four partial flags plus "11", codon_start on each strand, /inference and /anticodon rules, qualifier filtering and ordering, multi-contig grouping, and that the writer doesn't mutate the SeqFeatures it shares with the GenBank writer. One test asserts a plus-strand partial and its reverse complement produce mirrored output.

Full suite: 85 passed, 18 skipped. Ruff clean on new files; unchanged counts on the files touched.

End-to-end on both input routes: SAOMS1.gbk (246 CDS, 78 reverse strand, 3 tRNAs) and a truncated NC_043029 through pyrodigal-gv in both orientations — zero misoriented coordinate lines, forward/revcomp producing exact mirrors.

docs/output.md updated.