[m-murphy]

Summary

Sampler changes that mix the transmission network faster. Every sampler targets the same posterior as before.

 Sampler set (per Infection)
   RandomAllelesBitSetSampler3   # parent-set genotype proposal
+    parent-set option drawn from the parent-set posterior (10% uniform), four loci per redraw
   ParentRelinkSampler
+    candidate parents weighted by observed-allele coverage
   StrainCountSampler
+    observed-allele toggles
   CloneOrderSwapSampler
+    clone infection-time exchange
 Sampler set (shared)
   SALTSampler (allele frequencies)
+    self-fitting Dirichlet independence move in blocks of 4 alleles; fits freeze when adaptation stops
+  GuidedPairSampler              # new: time exchange for mutual pairs, deterministic pilot, pairs picked uniformly

Tuning changes:

• The genotype proposals (RandomAllelesBitSetSampler3 and 4, ParentRelinkSampler) read their per-allele rates from the Infection's sampled expected false positives and negatives, not fixed 0.01 rates.
• RELINK_TRIES 2 → 1: ParentRelinkSampler's multiple-try move is off.
• INFORMED_JUMP_MAX_TRIES 8 → 4.
• SALTSampler updates every Sampling iteration, with a full SALT sweep every 30th update and the Dirichlet move every 5th.
• RandomAllelesBitSetSampler4 updates every 10th iteration at loci with an observed allele (every iteration at unobserved loci), and a path that changes nothing is no longer counted as a proposal.

Other changes:

• AbstractSampler::countProposal counts one acceptance outcome for an update built from several Proposals.
• The glossary gains Mutual pair, Clone, Pilot and Dirichlet independence move.

Evidence

• Before: these proposals didn't exist.
  After: engine tests 311 pass (1 skipped), among them exact-draws invariance tests for GuidedPairSampler, the parent-set proposal with state-dependent option weights and the one-locus redraw, and a test of SALTSampler's Dirichlet move alone that fails without its Hastings correction. plasmotrackr tests 3724 pass.
• Golden Stores re-recorded: a seed now gives different draws, and the Store gains GuidedPairSampler's acceptance rows. plasmotrackr's hand sums are updated to the new rows.
• The pilot is deterministic only up to rounding. It reads its gains from the target's running sums and compares them with each other and with a 1e-9 threshold. When two gains tie in exact arithmetic, rounding can send the reverse pilot down another path, and that swap's Hastings ratio is then slightly off.

Merge Danger

Door: two-way

Blast Radius: Runs

• A seed gives different draws than before.
• A checkpoint written by an older build has a different sampler list, so it is refused on --resume or --start-from.