[m-murphy]

Summary

Family moves join every Sampler set except the null model's, which has no transmission. A family is a close mutual pair, or a connected triple of close pairs. A pair is close when at most 3 loci have an allele observed in one infection and not the other, counted in the direction where that count is smaller. Loci missing in either infection don't count.

 Sampler set
   … every sampler before
   GuidedPairSampler/pairs
+  GuidedGroupSampler/families      # 5 relocations per iteration: two members to new infection times,
+                                   # genotypes from a pilot over parents inside the family, exact reverse check
+  PairLocusGibbsSampler/families   # 20 Proposals per iteration: a pair's genotypes at one Locus redrawn together

No option: like every other sampler, they're always on outside the null model. The counts are FAMILY_RELOCATIONS and FAMILY_PAIR_LOCUS_PROPOSALS in config.h, the one source of the family moves' settings. Both samplers use one closeness test (Family.h), and the three guided samplers share one log-weighted draw. CONTEXT.md gains Family and Family move.

Evidence

• Before / after, six simulated studies × 4 chains, 3000 iterations, sampled log-uniform leak, mean over studies:

  	R-hat	chain spread	edges with chains > 0.5 apart	edge AP	top-1	wall time
  without family moves	1.05	11.7	2.4	0.658	0.502	56.6 h
  with family moves	1.03	6.2	1.1	0.666	0.506	65.8 h

  Chain disagreement roughly halves, for about 16% more time. In an earlier run, 50 relocations and 200 pair-locus Proposals cost about twice as much as 5 and 20, with no gain.

• FamilyMovesTest checks both moves leave their target invariant against exact draws, for a pair and for a relocation of a triple. A third test checks that missing loci don't count toward closeness. Engine tests: 316 pass, 1 skipped. plasmotrackr tests: 3724 pass.

• The example input has no families, so the golden Stores keep every draw. They gain only the two samplers' rows in structure/samplers and in acceptance.

Merge Danger

Door: two-way

Blast Radius: Runs

• Inputs with families give different draws for a seed and run about 16% longer.
• A checkpoint from an older build is refused (different sampler list).