test: withr-free local seed and RNG-state leak fixes

[schochastics]

What & why

The test suite has an opt-in check (gated behind IGRAPH_CHECK_RNG_STATE=true in
tests/testthat/setup.R) that fails any test which leaves the global
.Random.seed changed. This PR makes the whole suite pass that check and moves
the suite's RNG-seed handling off withr onto in-house helpers.

Running the suite with IGRAPH_CHECK_RNG_STATE=true: 0 seed leaks, 0
failures, 0 warnings (9739 passing).

New helpers (tests/testthat/helper.R)

igraph_local_seed() — drop-in for withr::local_seed(). Sets the seed for
the current scope and restores the previous global RNG state when that scope
exits.

• Self-contained: snapshots and restores RNGkind + .Random.seed itself
  (uses withr::defer only as the deferral primitive), so it doesn't inherit
  withr::local_seed's past seed-handling bugs.
• Restores RNGkind first, then .Random.seed last — changing the kind
  discards the seed, so the seed write must come last for an exact round-trip.
• Optional rng_version (e.g. "3.5.0"), folding in what used to be a separate
  local_rng_version() call. Pairing two independent deferred restores (kind +
  seed) ran them in the wrong order and leaked; a single ordered restore fixes it.

igraph_with_seed() — block form mirroring withr::with_seed(): set the
seed, evaluate the block, restore. Needed where a single test has several
independently seeded blocks with code between them (e.g. test-print-classic.R),
which a frame-scoped igraph_local_seed() can't express.

setup.R

• Set a baseline seed (set.seed(42)) so .Random.seed always exists before
  tests run. Otherwise the first RNG-touching test creates it from nothing
  (many igraph C functions initialise the RNG without drawing from it, e.g.
  voronoi_cells()), which the check flagged — and restoring that absent state
  just shoved the "first creation" onto the next test (whack-a-mole).
• The seed-leak override now forwards the test block via
  substitute() + eval(as.call(...)) instead of
  rlang::inject(testthat::test_that(name, !!code)). inject() eagerly
  processed any injection operator in the test body (!!, !!!, {{),
  e.g. set_vertex_attrs(g, !!!attr_list), evaluating it in the wrong frame and
  erroring with object 'attr_list' not found.
• The check stays gated behind IGRAPH_CHECK_RNG_STATE=true (opt-in for CI).

Test files

• Moved helper functions that were defined inside test files into a shared
  helper-test-functions.R (the override re-scopes test code, so file-local
  helpers were no longer reachable from the test body).
• Added igraph_local_seed() to tests that genuinely advance the RNG (unseeded
  sample_* fixtures and ARPACK eigensolvers).
• Replaced all 875 withr::local_seed() call sites with
  igraph_local_seed() (including the 732 in test-aaa-auto.R).
• Replaced the 9 withr::with_seed() call sites with igraph_with_seed()
  (test-print-classic.R, test-layout.R, test-iterators.R). These never
  leaked — consistency cleanup; snapshots unchanged.
• Replaced every local_rng_version("3.5.0") + seed pairing with
  igraph_local_seed(seed, rng_version = "3.5.0"), then removed the now-unused
  local_rng_version() helper.
• restore_rng_state() suppresses warnings around the RNGkind restore so
  putting back the older "Rounding" sampler (RNGversion("3.5.0")) doesn't
  emit R's "non-uniform 'Rounding' sampler used" note.

tools/AGENTS.md

Updated the test-writing guidance to use igraph_local_seed, so newly added
test-aaa-auto.R tests follow the convention.

Notes

• igraph_local_seed()/igraph_with_seed() live in helper.R (not setup.R)
  so helper functions like make_scan_graphs() can see them.
• The suite no longer uses withr for any RNG-seed handling.

🤖 Generated with Claude Code