Pipeline diagnostics

• rescue_redundancy [TRIGGERED]: 15124757/19522195 (77.5%) rescuable cells overlap a predicted gene already assigned to a different family in that strain -- the same locus double-counted, not a genuine annotation dropout (issue #133).
  • prune option: do not enable rescue (--pangenome_rescue_enable false); see clustering sweep #132 before re-enabling
  • prune option: verify --pangenome_rescue_structural_filter is enabled (default true) -- it is the mechanism that removes these cells, not merely detects them
• assembly_quality_confound [TRIGGERED]: max |rho| = 0.41 (rho = 0.4104, accessory_present vs n_contigs (partial, controlling total_length), n = 360 strains); threshold 0.3. See assembly_quality_report.md (issue #130). Between-strain accessory-content comparisons are confounded by assembly fragmentation on this dataset.
  • prune option: exclude or flag the most fragmented assemblies (highest n_contigs in assembly_quality_vs_content.tsv) and rerun
  • prune option: report accessory-content comparisons with assembly quality as a covariate, not as raw per-strain counts
• clade_structure_validity [not computed]: trips when: PCoA1 variance share < ~50%, split-half ARI < 0.8, or between/within clade distance ratio < ~2. Not yet computed by this pipeline -- see #131.
• gain_loss_skew [not computed]: trips when: gain:loss ratio > 20:1 or ambiguous-polarity fraction < 2%. Not yet computed by this pipeline -- see issue #134's calibration reference.