[ms609-agent]

Research prototype for #174, with no package code changes. It asks whether a Pol & Escapa (2009)-style leaf-instability score can choose which leaf SPRDist(method = "Rogue") drops, in place of the per-candidate bounded-MCI scoring.

Write-up: dev/rogue_polescapa_findings.md. Scripts, reference set and results: dev/rogue_polescapa/.

Method. Each split of one tree is treated as a binary character and scored with Fitch on the other tree, in both directions. A leaf's score is the number of steps saved when that leaf is coded ? in every character. One downpass and one uppass give every leaf's score. A bit-parallel C++ kernel is included (fitch_bitpar.cpp).

Recommendation. Adopt it as the selector, with MCI used only to break ties among the leaves with the top Fitch score (both-tied).

• Exact references. These are 528 pairs of ≤ 32 leaves with exact uSPR, including the two fixtures recovered from git history. both-tied is equal or better than the current selector everywhere except random pairs of 8–14 leaves: there it over-estimates by one move on 21 of 120 pairs, against 9 for the current selector.
• 50–150 random leaves. Measured against best-known (the lowest estimate any variant gave), the mean excess is 0.47–0.90 moves, against 1.97–4.60 for the current selector. It gives a lower estimate on 52 pairs and a higher one on 1.
• Cost. The Fitch kernel takes 25–170 µs per iteration, against 1.5–108 ms for spr_rogue_scores(). A harness that uses pure Fitch runs about 2×, 10× and 40× faster than SPRDist at 30, 80 and 150 leaves. The cost of both-tied is projected rather than measured: it would score 2–3 tied candidates in 30–46% of iterations.
• Variants that lose. One scoring direction alone does worse than the current selector. A top-k pre-filter followed by MCI does no better than both-tied and costs more.

A C++ implementation is being worked on separately and depends on the src/spr.cpp incremental-fill work landing first.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.74%. Comparing base (0dd26a8) to head (0214785).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #226      +/-   ##
==========================================
- Coverage   97.74%   97.74%   -0.01%     
==========================================
  Files          60       60              
  Lines        6577     6575       -2     
==========================================
- Hits         6429     6427       -2     
  Misses        148      148              

☔ 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.