[gouyt13]

Closes #216

HNSW-RaBitQ, SymphonyQG and the hnswlib graph behind IVF-RaBitQ's HNSW routing do not guarantee that every point is reachable. A multi-threaded or unseeded build can leave a point with no incoming edge, so tests that require every point to be found fail now and then in CI. #227 fixed two Python cases. This PR finishes the audit of the graph tests in tests/unit/rabitqlib/index/ and tests/python/.

Changes

Two approaches, depending on what a test checks.

• Exact checks → fixed seed and one thread.
  • test_large_dimensions.py: SymphonyQG is seeded and builds with one thread. HNSW already used one thread and the default seed.
  • test_symqg.py: the shared built_symqg fixture uses the 0.8 form with seed=42 and one thread. The rebuild case passes seed=42 to the 0.7 keyword form.
  • test_portable_indexes.py: the SymphonyQG case uses the 0.8 form with seed=42 and one thread.
  • test_hnsw.py: three cases pass num_threads=1 to add, where the Python default is every CPU:
    • the dense-labels case compares the first ten points' results before and after add;
    • the resize case requires those points afterwards;
    • the remove case compares graphs grown separately from two loaded copies.
• Parallel construction → keep the threads, check self-recall of every point.
  • test_examples.py: the example scripts build graphs with two threads, and SymphonyQG also has no fixed seed. HNSW and SymphonyQG now require a self-recall of at least 95% over all 128 points. IVF keeps its exact check.
  • test_ivf_initializers.py: the parallel HNSW-routing build over 1,024 centroids requires a self-recall of at least 99% over every point.

Tests that stay exact

Two tests listed in #216 keep their exact checks, although they build with several threads:

• HnswAddTest.ParallelAddKeepsLabelsAndTheGraph;
• the Python test_parallel_build_preserves_self_retrieval.

Both set M to the point count. The base layer therefore holds 2M ≥ N − 1 links, so mutually_connect never prunes a reverse link, and every point stays reachable under any thread interleaving. Only their comments change, to say this. The same holds for the dense-graph add test, which keeps the default thread count, and for hnswlib routing over 32 or 33 centroids (kM = 16, so maxM0_ = 32).

An aggregate check would have weakened these tests. An experimental build that drops 5 of 512 reverse links passed a 99% self-recall check, and failed the exact one.

Relation to #214

test_hnsw.py, test_symqg.py, test_portable_indexes.py and test_ivf_initializers.py are still on the #214 list. This PR changes only the cases above. In test_symqg.py and test_portable_indexes.py, the SymphonyQG fixture and the portable case already use the 0.8 form, so that they can be seeded.

Checks

All on Linux x86-64, Xeon Gold 6418H (96 CPUs), GCC 14.3, Release, RABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF; the Python package was built from this branch.

• Full suites: C++ 499/499; Python 1146 passed, 2 skipped. check-format.sh and check-python.sh pass.

• Loops of the changed and restored tests, all with 0 failures:

  Test	Runs
  HnswAddTest.ParallelAddKeepsLabelsAndTheGraph	1,000 on 96 CPUs, 1,000 restricted to 4 CPUs
  changed test_hnsw.py cases, test_symqg.py, test_portable_indexes.py, the parallel routing case, test_large_dimensions.py	1,008
  the two dense HNSW Python cases	200, restricted to 4 CPUs
  the HNSW and SymphonyQG example cases	208
  test_large_dimensions.py	1,000 more

• Still fails when broken: a search returning wrong IDs fails every changed check. The exact checks also catch a lost reverse link, as described above. Apart from that experiment, the Python changes were not mutation-tested.

• Review: an independent review found that a first version had weakened the two tests above, and had changed routing tests that cannot fail. Those changes were reverted.

Not covered

• The SymphonyQG alignment failure reported in Make graph-index tests independent of graph connectivity #216 is not explained. It failed about 0.2% of the time inside pytest, and never in separate processes. test(python): migrate ten test files off the 0.7 call forms #227 removed the assertion that failed, so it no longer affects CI.
• Other platforms: a seeded single-thread build is reproducible on one platform, not necessarily across platforms. CI covers arm64, macOS and Windows.