[glennhickey]

Two changes to the multiple aligner, both for cactus's bar. ComparativeGenomicsToolkit/cactus#2024 pins this branch.

Per-pair state machines (ccdbc32).

• makeAlignmentWithPairStateMachines takes a callback that picks the state machine for each pair of sequences it aligns. That lets one multiple alignment score different pairs with different models; in cactus, by how diverged the pair's genomes are.
• A NULL callback, or a NULL from it, means the default machine. makeAlignment, makeAlignmentUsingAllPairs and makeAllPairwiseAlignments keep their signatures and behaviour.
• The new test checks three things: a chooser returning the default changes nothing; it's asked once per pairwise alignment, about a real pair; and another machine gives a valid but different alignment.

Deterministic multiple alignment (2ec01d1). The same sequences, models and parameters now always give the same alignment. Before, the result could depend on:

• sonLib's random number generator, drawn from for each aligned pair's tie-breaking jitter, and for a coin flip between equally good pairs when choosing spanning-tree alignments. So the result depended on whatever had drawn from it before.
• Where memory had been allocated. Weights with equal scores were ordered by address, and which of a weight and its reverse went into the ordered set was decided by comparing column addresses.

Now the jitter is a hash of the aligned pair, with the same magnitude and the same value for both directions. Ties go by the columns' positions, and the spanning-tree tie-break goes by a hash of the pair of sequences.

Testing:

• Old behaviour: in 400 random cases, two identical calls of the old code gave different aligned pairs in 16 of the 201 that took the spanning-tree path, and none of the 199 all-pairs cases.
• New tests: the tests now require identical output (aligned pairs, chosen pairwise alignments and columns, in order). A new test advances the generator and holds extra memory between two calls. The old code fails both under each of 24 seeds; this passes them all. All 42 tests pass.
• In cactus: on evolver mammals with pecan, two runs are identical on every genome. Against faster-pecan, 17 of 4.5 million leaf-pair alignments changed, with accuracy the same to four decimals.

🤖 Generated with Claude Code