perf(ZIBR): single-pass relabel-and-sum for pure bus merges

[luke-kiernan]

Stacked on #329. Removes the Build-Ybus regression that #329's correctness fix surfaced.

Why

#329's per-branch L2 criterion correctly merges low-impedance branches that the old iszero(real(Y)) gate skipped. On main these large systems merged 0 buses, so the reduction short-circuited (isempty(reduction) && return ybus); now they merge a handful (28 on the Eastern Interconnect), which makes the merge machinery hot for the first time.

That machinery — _merge_ybus_buses! → _accumulate_csc_col_into! — sums each removed bus's row/column into its survivor with per-entry CSC structural inserts, each of which shifts the whole backing store of a 78k×78k matrix, then slices out the removed buses with data[bus_ix, bus_ix]. Cost scales with merges × nnz.

Measured (Eastern, 78,484 buses, same machine):

	Build Ybus	merged
main	~313 ms	0
#329 (slow merge)	~505 ms	28
this branch	~375 ms	28

Residual merge overhead over the no-merge baseline drops from ~140–190 ms to ~40 ms.

What

A pure bus merge (only ZIBR produces one — radial/Ward add admittances, degree-two adds series branches) is just an index relabeling, so the merge and the bus-removal slice are fused into a single O(nnz) pass: relabel each entry's row/column to the surviving index and let sparse() sum collisions. Adjacency is rebuilt the same way, then re-derived for survivors (anti-parallel branches cancel in the signed sum, mirroring the existing repair).

• Guarded by _is_pure_merge_reduction (merges present; no removed_buses/series_branch_map/removed_arc_to_surviving_bus/added maps). The general elimination path is untouched, so radial/degree-two/Ward are byte-for-byte unchanged.
• Each reduction is its own _apply_reduction call, so the fast path fires on the ZIBR step even when other reductions follow it.

Tests

• New ZIBR fast merge matches a rewire-and-rebuild oracle: merging bus 3 into bus 2 across a zero-impedance line equals an independently built system where bus 3's branch is rewired onto bus 2 and the ZI line removed (the huge ZI entries cancel in the row/column sum).
• Full reduction suite (ZIBR, anti-parallel, radial, degree-two, Ward, composition, 14-bus combos) — 33k assertions — passes, including the adjacency-sensitive anti-parallel merge keeps a bus's degree.

🤖 Generated with Claude Code

[github-actions]

Performance Results

Precompile Time

Main	This Branch	Delta
2.2 s	2.1 s	-2.1%

Execution Time

Cells show median (min–max) over 5 samples; delta compares medians.

Test	Main	This Branch	Delta
matpower_ACTIVSg2000_sys-Build Ybus	9.1 (9.0–11.3) ms	9.9 (9.8–12.6) ms	+9.2%
matpower_ACTIVSg2000_sys-Build PTDF	72.1 (71.5–72.5) ms	74.0 (72.7–77.8) ms	+2.7%
matpower_ACTIVSg2000_sys-Build LODF	145.9 (144.6–149.7) ms	151.0 (148.9–152.5) ms	+3.5%
matpower_ACTIVSg2000_sys-Build VirtualMODF	90.5 (90.3–94.8) ms	1.0 (1.0–1.0) s	+1024.5%
matpower_ACTIVSg2000_sys-VirtualMODF Query 10 rows	0.6 (0.5–1.3) ms	0.7 (0.7–1.6) ms	+15.2%
matpower_ACTIVSg2000_sys-Radial network reduction	0.7 (0.7–1.0) ms	0.9 (0.9–1.2) ms	+22.6%
matpower_ACTIVSg2000_sys-Degree two network reduction	0.9 (0.9–1.4) ms	1.0 (1.0–1.6) ms	+2.1%
Base_Eastern_Interconnect_515GW-Build Ybus	526.2 (522.8–555.2) ms	607.3 (604.9–626.6) ms	+15.4%
Base_Eastern_Interconnect_515GW-Build VirtualPTDF	762.9 (761.2–765.0) ms	834.8 (832.2–840.5) ms	+9.4%
Base_Eastern_Interconnect_515GW-VirtualPTDF Query 10 rows	24.2 (23.9–24.3) ms	25.3 (24.9–25.5) ms	+4.3%
Base_Eastern_Interconnect_515GW-Build VirtualMODF	759.9 (754.5–761.3) ms	922.6 (918.9–922.9) ms	+21.4%
Base_Eastern_Interconnect_515GW-VirtualMODF Query 10 rows	28.7 (28.3–29.0) ms	29.5 (29.1–30.1) ms	+2.5%
Base_Eastern_Interconnect_515GW-Radial network reduction	33.1 (32.9–34.2) ms	33.3 (32.9–33.8) ms	+0.5%
Base_Eastern_Interconnect_515GW-Degree two network reduction	47.8 (47.4–48.8) ms	50.0 (47.6–50.9) ms	+4.5%