[krlmlr]

The tests of the generated *_impl() functions move from one 11,000-line test-aaa-auto.R into one file per category, mirroring the R/aaa-<category>.R files that define the impls, and a script keeps them there. It is stacked on #2999.

What it changes

• The file-level note at the top of test-aaa-auto.R goes. It asked for snapshot and structured tests for each impl, which the "Testing Pattern" section of tools/AGENTS.md already says, and it would have no file to live in after the split.
• test-aaa-auto.R is split into 26 files test-aaa-<category>.R, and _snaps/aaa-auto.md into the matching _snaps/aaa-<category>.md. This commit is a pure move generated by the script below: each test_that() block, with the comments above it, and each snapshot section moves byte for byte, in the original order, and every file keeps the skip_on_cran() header of the old one.
• tools/split-aaa-tests.R puts each impl's tests in the file of the category R/aaa-<category>.R assigns it, so a rerun after an impl changes category in tools/aaa-categories.yaml moves its tests along. With --check it writes nothing and fails when a rerun would change a file; CI runs it so in the after-install step. Top-level code other than test blocks is an error, because it would have no file to move to.
• The handbook (testing/suite/, architecture/r-layer/new-functions/) and tools/AGENTS.md name the per-category files.

Checks

• Rscript tools/split-aaa-tests.R --check reports all 729 impl tests in place, and a rerun without --check writes nothing.
• Comparing sorted lines, the 26 new test files hold exactly the lines of the old file plus the two-line header repeated per file, and the snapshot files hold exactly the lines of the old one.
• All 729 blocks pass under testthat::test_local(). Before the second commit of fix: Calling igraph from a warning handler no longer corrupts the call that warned #2999 they did not: six snapshots lost the call of their warning under load_all(), which is how that problem was found.
• devtools::document() and air format change nothing, and the handbook check script is clean.

Config/testthat/start-first in DESCRIPTION now names aaa-* in place of aaa-auto, a glob testthat accepts, so the split files still start first.

Starting the split files first changed the RNG state that each test worker hands on to later files, and that exposed an unseeded pin from #2997: eigen_centrality() on make_ring(3) + make_ring(4), whose two rings share the leading eigenvalue, returns seven ones or one ring of zeros depending on the random perturbation of ARPACK's start vector, the zeros for about four seeds in ten. The coverage run of the smoke test hit the zeros. The pin now calls igraph_local_seed(42), like the other seeded tests in test-centrality.R, and so do the other #2997 pins that draw random numbers, hits_scores() on a ring and cluster_infomap() on the karate graph, whose results do not depend on the seed but which left the RNG advanced for the files after them, against the rule in setup.R.

LLM use: this PR was prepared with Claude Code (Claude Opus 5.5), at the maintainer's request. The change comes from the subsession https://claude.ai/code/session_01Dmrf8jeMGkMHFX2nqRKXfD.

• By submitting this pull request, I assign the copyright of my contribution to The igraph development team.

---

Generated by Claude Code

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 8a884a2 is merged into claude/v2-warning-reentrancy:

• ✔️E_named: 228ms -> 228ms [-1.73%, +1.14%]
• ✔️V_named: 434ms -> 432ms [-0.89%, +0.15%]
• ✔️all_simple_paths_named: 149ms -> 150ms [-0.51%, +1.55%]
• ✔️as_adjacency_matrix: 225ms -> 225ms [-0.66%, +0.35%]
• ✔️as_biadjacency_matrix: 255ms -> 255ms [-0.37%, +0.32%]
• ✔️as_data_frame_both: 245ms -> 244ms [-1.57%, +0.38%]
• ✔️as_long_data_frame: 213ms -> 214ms [-0.52%, +0.75%]
• ✔️degree_few_vertices: 127ms -> 127ms [-0.32%, +1.47%]
• ✔️ego_order2_named: 157ms -> 157ms [-0.97%, +0.96%]
• ✔️es_attr_assign: 238ms -> 239ms [-0.6%, +0.87%]
• ✔️es_attr_filter: 245ms -> 244ms [-1.33%, +0.64%]
• ✔️graph_from_adjacency_matrix: 268ms -> 267ms [-1.01%, +0.24%]
• ✔️graph_from_data_frame: 284ms -> 284ms [-0.62%, +0.6%]
• ✔️head_of_named: 180ms -> 180ms [-0.97%, +0.95%]
• ✔️incident_edges: 214ms -> 214ms [-0.96%, +1.01%]
• ✔️max_cliques_named: 26.5ms -> 26.3ms [-1.06%, +0.04%]
• ✔️neighbors_by_name: 244ms -> 244ms [-0.98%, +0.64%]
• ✔️union_named: 234ms -> 233ms [-1.25%, +0.76%]
• ✔️vs_attr_filter: 329ms -> 328ms [-0.97%, +0.3%]
• ✔️vs_by_name: 305ms -> 303ms [-1.51%, +0.55%]
  Further explanation regarding interpretation and methodology can be found in the documentation.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if f9de58a is merged into claude/v2-warning-reentrancy:

• ✔️E_named: 230ms -> 228ms [-2.66%, +1.18%]
• ✔️V_named: 432ms -> 431ms [-0.74%, +0.26%]
• ✔️all_simple_paths_named: 149ms -> 149ms [-0.86%, +0.72%]
• ✔️as_adjacency_matrix: 230ms -> 230ms [-0.79%, +0.89%]
• ✔️as_biadjacency_matrix: 260ms -> 260ms [-0.53%, +1.22%]
• ✔️as_data_frame_both: 246ms -> 246ms [-0.66%, +0.48%]
• ✔️as_long_data_frame: 219ms -> 219ms [-1.11%, +1.29%]
• ✔️degree_few_vertices: 127ms -> 127ms [-1.02%, +0.85%]
• ✔️ego_order2_named: 157ms -> 157ms [-1.01%, +0.8%]
• ✔️es_attr_assign: 244ms -> 243ms [-1.8%, +1.01%]
• ✔️es_attr_filter: 251ms -> 252ms [-1.22%, +1.7%]
• ✔️graph_from_adjacency_matrix: 272ms -> 272ms [-0.82%, +0.44%]
• ✔️graph_from_data_frame: 288ms -> 289ms [-0.56%, +1.23%]
• ✔️head_of_named: 181ms -> 181ms [-0.73%, +1.32%]
• ✔️incident_edges: 217ms -> 217ms [-0.68%, +1.03%]
• ✔️max_cliques_named: 26.8ms -> 26.7ms [-1.03%, +0.5%]
• ✔️neighbors_by_name: 247ms -> 247ms [-0.71%, +0.36%]
• ✔️union_named: 249ms -> 248ms [-3.43%, +2.86%]
• ✔️vs_attr_filter: 335ms -> 334ms [-0.91%, +0.13%]
• ✔️vs_by_name: 311ms -> 311ms [-0.98%, +1.02%]
  Further explanation regarding interpretation and methodology can be found in the documentation.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 552eccb is merged into claude/v2-warning-reentrancy:

• ✔️E_named: 219ms -> 217ms [-2.09%, +0.33%]
• ✔️V_named: 438ms -> 439ms [-0.34%, +0.76%]
• ✔️all_simple_paths_named: 150ms -> 150ms [-0.29%, +1.18%]
• ✔️as_adjacency_matrix: 228ms -> 229ms [-0.31%, +0.63%]
• ✔️as_biadjacency_matrix: 259ms -> 259ms [-0.52%, +0.54%]
• ✔️as_data_frame_both: 244ms -> 244ms [-0.83%, +0.26%]
• ✔️as_long_data_frame: 215ms -> 214ms [-0.8%, +0.37%]
• ✔️degree_few_vertices: 128ms -> 127ms [-0.95%, +0.5%]
• ✔️ego_order2_named: 157ms -> 157ms [-1.07%, +0.62%]
• ✔️es_attr_assign: 239ms -> 238ms [-1.32%, +0.34%]
• ✔️es_attr_filter: 246ms -> 245ms [-1.03%, +0.55%]
• ✔️graph_from_adjacency_matrix: 272ms -> 272ms [-0.52%, +0.63%]
• ✔️graph_from_data_frame: 284ms -> 285ms [-0.4%, +0.7%]
• ✔️head_of_named: 180ms -> 181ms [-0.33%, +1.64%]
• ✔️incident_edges: 214ms -> 213ms [-1.09%, +0.48%]
• ✔️max_cliques_named: 26.4ms -> 26.4ms [-0.36%, +0.5%]
• ✔️neighbors_by_name: 245ms -> 244ms [-0.84%, +0.54%]
• ✔️union_named: 234ms -> 236ms [-0.82%, +2.03%]
• ✔️vs_attr_filter: 332ms -> 331ms [-0.68%, +0.49%]
• ✔️vs_by_name: 305ms -> 305ms [-0.68%, +0.85%]
  Further explanation regarding interpretation and methodology can be found in the documentation.