Bound Rogue SPR candidates with parent dual prices

[ms609]

Before: for almost every candidate leaf, SPRDist(method = "Rogue") solves a full LAP, and 89% of those LAPs end below the running minimum. The LAP accounts for 52% of runtime at 80 leaves and 66% at 150.

After: the parent pair's LAP is solved once, lazily, on the first candidate that passes the row bound. Each candidate's splits are then priced with the parent's column duals.

• For any prices v, LB(v) = Σ_j v_j + Σ_i min_j (c_ij − v_j) is at most the optimal assignment cost. The bound is therefore exact in int64, and the prices affect only speed.
• The bound rejects 82–88% of the LAPs that would otherwise only reject.
• The remaining LAPs warm-start JV from v (src/lap_warm.cpp, kept in its own translation unit so that lap()'s codegen is untouched).

Wall-clock (prototype, eager prices; minimum over interleaved runs):

Workload	Baseline (s)	This PR (s)	Change
30 leaves, random	0.423	0.343	−19%
80 leaves, random	4.00	1.94	−51%
150 leaves, random	18.2	6.40	−65%
150 leaves, 30 moves apart	1.31	1.00	−23%
Near-identical pairs			no change (noise)

Verification:

• SPRDist() output was identical on 345 pair runs.
• Across 650,676 candidate scores the largest difference was 7e-15, and the argmin sets were identical.
• After this rebase onto main, SPRDist() matches the R path on 40 random pairs of 8–60 leaves.
• Local covr on the SPR tests leaves no new line uncovered. The unreachable guards are marked #nocov.

The research write-up and scripts are in dev/rogue_mci_bound_findings.md and dev/rogue_bound/, which .Rbuildignore excludes from the package build.

🤖 Generated with Claude Code

https://claude.ai/code/session_014CDmnvPhnKFWMrQgqPKVY7

[github-actions]

Performance benchmark results

Call	Status	Change	Time (ms)
ClusteringInfoDistance(tr280)	⚪ NSD	0.19%	9.29, 9.26 → 9.27, 9.22
ClusteringInfoDistance(tr50)	⚪ NSD	-0.69%	6.64, 6.63 → 6.73, 6.66
LAPJV(test2000)	⚪ NSD	-0.1%	95.3, 95.1 → 94.9, 95.3
LAPJV(test40)	⚪ NSD	-1.93%	0.0176, 0.0178 → 0.018, 0.0181
LAPJV(test400)	⚪ NSD	-0.91%	3.2, 3.21 → 3.23, 3.24
MutualClusteringInfo(tr280)	⚪ NSD	-1.21%	17.9, 18.2 → 18.1, 18.4
MutualClusteringInfo(tr50)	⚪ NSD	-0.19%	17.3, 17.2 → 17.8, 17.1
NNIMaxStep(2000, normalize = TRUE)	⚪ NSD	0.25%	0.215, 0.216 → 0.214, 0.216
NNIMaxStep(300, normalize = TRUE)	⚪ NSD	0%	0.0315, 0.0316 → 0.0317, 0.0315
NNIMaxStep(300)	⚪ NSD	-0.35%	0.0087, 0.00889 → 0.00886, 0.0088
PathDist(postTrees)	⚪ NSD	-0.27%	2.39, 2.38 → 2.39, 2.4
PhylogeneticInfoDistance(tr180)	⚪ NSD	1.01%	59, 57.9 → 58.3, 58.2
PhylogeneticInfoDistance(tr50)	⚪ NSD	0.3%	44.9, 43.5 → 43.7, 44.2
RobinsonFoulds(tr280)	⚪ NSD	0.21%	4.76, 4.75 → 4.75, 4.74
RobinsonFoulds(tr280)	⚪ NSD	0.11%	4.57, 4.57 → 4.6, 4.56
RobinsonFoulds(tr50)	⚪ NSD	-0.98%	4.57, 4.56 → 4.61, 4.59

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.57%. Comparing base (64e9127) to head (2095b68).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #215      +/-   ##
==========================================
+ Coverage   97.44%   97.57%   +0.12%     
==========================================
  Files          59       60       +1     
  Lines        5762     6065     +303     
==========================================
+ Hits         5615     5918     +303     
  Misses        147      147              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.