[hyphaltip]

Replaces the 3-point near-tie rule of #214 (curator decision, option A, 2026-10-29), after an independent evaluation of the margin (results/2026-10-29_near_tie/README.md).

What changes (all behind --cross-family-policy on)

• For groups of overlapping cross-family calls where neither family is scoped to the genome, the family is decided once per genome: the higher mean best core-gene identity over those groups wins every group (equal: smaller family name). No margin, no MATsc preference. A scoped family still wins its group.
• Every such kept call carries family_ambiguous: identities of both families, gap, genome means and locus counts (YAML block; GFF3 family_ambiguous=family:identity|...; new DetectionResult.family_ambiguous).
• family_unconfirmed is now only the PAP1/OBP1/PIK1 flank rule; cross_call_near_tie and NEAR_TIE_POINTS are gone.
• The set-aside call's gene evidence is kept in the cross_call note (gene, role, identity, coverage, status) so the two families can be compared gene for gene after a run.
• Tests rewritten (genome-level pooling, no margin, scoped group, third family, flag content). Suite: 1596 passed (draw/synteny tests need pygenomeviz, absent locally).

Evidence

Opus-class independent evaluation of the 156 cross-called pairs: no natural break at 3 points (the cutoff splits congeners), one residue is about 1.7 points so the margin is under two residues, the MATsc tie-break is biased (sla2 is searched only by MATsc, PAP1/OBP1/PIK1 only by MTL), per-locus decisions gave 4 mixed-family genomes. Replay of the 354 yeast genomes: mixed-family genomes 4 -> 0, near-tie flags 42 -> 0, 156 family_ambiguous loci in 140 genomes, kept-lower-identity 12 -> 1, 14 genomes change calls (family and idiomorph names), 14 more mtl_outside_scope flags.

Open for the curator

• No ground truth exists for these genomes; the family remains a best-hit identity choice. Per-genus rulings, profile scoring against both reference sets and synteny would settle it (README).
• Release prep (Release prep v0.8.0 (draft; do not merge until the release is cut) #207) and Labels: registry_version, frozen codes, label_status and the explorer marker (#210, #211) #217's label_status doc mention the old rule or family_unconfirmed; I will update the Release prep v0.8.0 (draft; do not merge until the release is cut) #207 CHANGELOG line.
• The replay database is main (no MATpk).

🤖 Generated with Claude Code