[gouyt13]

Summary

Migrate ten Python test files off the deprecated 0.7 call forms, so they pass under -W error::DeprecationWarning and keep working after #213 removes those forms in 0.9:

test_array_alignment.py, test_build_cluster_ids.py, test_concurrent_indexes.py, test_index_repr.py, test_kmeans_sampling.py, test_large_dimensions.py, test_persistence_paths.py, test_quantization_factors.py, test_remove_arrays.py, test_thread_counts.py

Related to #214 and #216. These are the remaining files from that issue that #222 does not touch. The other seven (test_hnsw, test_import, test_input_boundaries, test_ivf, test_ivf_initializers, test_portable_indexes, test_symqg) follow separately, to avoid conflicts with #222.

Changes

Most changes replace one call form with another and keep the same thread counts, nbits, seeds and ef_construction:

• 0.7 constructors become 0.8 keyword constructors. HNSW passes num_clusters=len(centroids) so that set_centroids accepts the centroids, and SymphonyQG uses nbits=32 instead of quantization_bits=0.
• build() becomes set_centroids() + add(cluster_ids=..., num_threads=1), or add(num_threads=...) for SymphonyQG.
• search_batch() becomes search(), and is_built becomes len().

Three files needed more than a direct replacement:

• test_concurrent_indexes.py used build() as a mutation that can run more than once, and 0.8 has no equivalent. Active searches and mutations are now checked to reject add, train and set_centroids with BUSY. No other test checked train or set_centroids for BUSY. These calls raise a different error when the index is idle, so a broken guard still fails the test. To show that SymphonyQG released its access, the test expects the "does not support incremental addition" error from a second add, which takes write access before it rejects the existing graph.
• test_array_alignment.py: IVF reproduces its rotation with seed=42, as HNSW already did. It no longer rebuilds a loaded reference, because 0.8 rejects set_centroids on an index that stores points. Aligned and unaligned inputs still save identical bytes.
• test_index_repr.py expects 0.8 capacities: IVF starts at 0, and HNSW calls resize(96) to keep a preallocated capacity. HNSW now fixes num_clusters=1 at construction, so the test no longer checks that the summary changes from the default 16 to 1.

#216 fixes in the two graph tests

CI on this branch failed twice. Each failure was a graph build that left one point unreachable, a case #216 already lists: first the qg alignment case, then test_thread_counts[18446744073709551615-l2-hnsw], where point 4's nearest result was 69. Both tests are in this PR, so the second commit applies the fixes proposed in #216:

• test_array_alignment.py: SymphonyQG now builds with seed=42 and one thread, like IVF and HNSW. Aligned and unaligned inputs must save identical bytes and return identical k=5 results; the qg case did not compare bytes before. The k=len(data) search, which required every point to be reachable, is removed. The exact-distance check now runs on the k=5 reference results.
• test_thread_counts.py: IVF still requires every point to find itself. HNSW and SymphonyQG require a self-recall@1 of at least 95%, so up to 4 of the 96 points may miss. Searches on the same graph must still agree across thread counts. The threshold is a judgment call: the only miss observed was the one in CI. Locally, 6,000 graph builds on 96 CPUs and 8,000 parallel builds restricted to 4 CPUs had no misses.

The migration did not change how the HNSW graph is built: add on an empty index runs the same link_exact as the old build (through construct), with the same seed, nbits, M and ef_construction.

Testing

Run with the package built from main; this PR changes only tests.

• ./scripts/check-python.sh and git diff --check pass.
• The ten files pass with -W error::DeprecationWarning: 191 passed.
• Full Python suite: 1146 passed, 2 skipped, the same as before the change.
• With -W error::DeprecationWarning on the full suite, only the seven files listed above still fail.
• test_concurrent_indexes.py passed 200 repeated runs, and test_array_alignment.py with test_thread_counts.py passed 200 repeated runs after the Make graph-index tests independent of graph connectivity #216 fixes.
• Before those fixes, the qg alignment case failed in about 0.2% of runs, on main as well (2 of 400 runs); details are in Make graph-index tests independent of graph connectivity #216.