[ms609-agent]

Follow-up to #223: extends the parent-overlap route from single-leaf to depth-2 pair Rogue candidates.

Change

• SingleDrop → LeafDrop, with Prepare(i, j). Two removals are replayed on copies of the parent ledgers (leaf i once per outer loop, then j per candidate), recording the split that vanishes from each tree at each step. This covers cherries, sib(i) = par(j), the mirror case and root-child transitions without enumerating them. The root-child test is now structural (parent[parent[tip]] == parent[1]): after removing i, sib(i) can become the root's child while still owning a row.
• Child overlap = parent overlap − popcount(row_has & col_has) on 2-bit masks. mutual_clustering_bounded_counts() dispatches on SplitCounts::pair to a template, so the depth-1 kernel is unchanged.
• Shared-split test: parent split pairs whose polarized XOR is exactly two leaves are stored as sorted pair keys. A pair also falls back to ScoreWithout() if either leaf alone makes a shared split, if either leaf is leaf 1, or if the reduction predicate fires on the doubly-reduced ledgers.
• Prices: when the first vanishing split's merge target vanishes second, the chain resolves to the second's target.

Results (min of 3 interleaved runs vs main, Windows, 1 core)

workload	main	this PR
rand30, depth 2	3.86 s	1.65 s (−57%)
mid80, depth 2	5.47 s	3.10 s (−43%)
near80, depth 2	0.39 s	0.21 s (−46%)
rand30 / rand80, depth 1	0.22 / 1.03–1.09 s	0.21 / 1.07–1.11 s (no change beyond noise)

SPRDist sums are identical on all workloads. 75% of incremental pair candidates are scored from overlaps. The rest are leaf 1 or disabled trees (14%), single-leaf shared splits (7%), pair-shared splits (1.6%) and reductions (1%).

Verification

• 295 tree pairs (9–40 leaves, half near pairs), 70,142 pair candidates, built with -DDEBUG (ASSERTs on): argmin sets of single and pair score vectors are identical to incremental = FALSE. The largest difference among near-best scores is 7e-15.
• test-tree_distance_spr.R passes (its Compare() already checks depth 2 against the rebuild path). All new C++ lines are covered (covr). Spelling is clean.

Note: a -DDEBUG build fails in nni_distance.cpp:212 (n_splits is undeclared in an ASSERT). This predates the PR and is unrelated.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.76%. Comparing base (2f6776d) to head (6e0d5a3).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #224      +/-   ##
==========================================
+ Coverage   97.74%   97.76%   +0.01%     
==========================================
  Files          60       60              
  Lines        6575     6626      +51     
==========================================
+ Hits         6427     6478      +51     
  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.