[codex] Speed up ueco refinement

[DanielSeemaier]

Summary

This tunes the shared-memory -P ueco preset and trims some overhead from the unconstrained LP/FM refinement path.

Changes:

• reduce -P ueco unconstrained FM seed nodes from 25 to 50 while increasing FM iterations to 30
• replace hot unconstrained LP shared counter atomics with thread-local reductions
• avoid full-graph partition copies in unconstrained FM and restore only nodes touched by the round move log during rollback

Benchmark

Single-server mkexp2 run, as requested:

• Experiment: /nfs/work/seemaier/experiments/2026.05.23-codex-ueco-final-naur-k64
• Server: naur only
• Graph set: /nfs/work/graph_benchmark_sets/ufm_paper/small
• k=64, epsilon 0.03, seed 1, topology 1x1x64
• Baseline: main (b4e1afd351cfd96e7228bde48fda36f46b830b95)
• Codex: codex/speed-up-ueco (fe6f3a4ca67a0af15fe6cb56ac4e3a02b5bfd02a)

Results over 44 paired instances:

Metric	Baseline	Codex	Delta
Total partitioning time	325.148 s	254.995 s	1.275x speedup / 27.512% faster
Aggregate edge cut	725,440,448	732,222,303	+0.935%
Feasible partitions	44/44	44/44	unchanged

Quality notes:

• Worst per-graph cut delta: recomp_sources1GB_9, 16,719,986 -> 21,090,980 (+26.142%)
• Best per-graph cut delta: rhg, 107,269 -> 100,739 (-6.087%)
• The large recomp_sources1GB_9 cut regression is the main tradeoff hidden by the aggregate.

Timing notes:

• Biggest speedup: recomp_sources1GB_9, 5.410 s -> 2.571 s
• Next biggest speedup: soc-sinaweibo, 56.551 s -> 28.962 s
• Worst slowdown: channel-500x100x100-b050, 1.836 s -> 2.072 s

Validation

• Built KaMinParApp
• Ran and passed:
  • build/tests/test_shm_endtoend
  • build/tests/test_shm_metrics
  • build/tests/test_shm_gain_caches
• mkexp2 parse completed successfully for the naur-only experiment. mkexp2 stats could not run because the remote R environment is missing tidyverse, so the summary above was computed from mkexp2's parsed CSV files.