[gouyt13]

Summary

Breaking C++ API change. The query correction-factor accessors now use the g_ prefix, the same as g_add() and g_error(). The old names are removed, with no deprecated aliases.

Old	New	Classes
k1xsumq()	g_k1xsumq()	SplitSingleQuery, BatchQuery, SplitBatchQuery, symqg::QuantizedQuery
kbxsumq()	g_kbxsumq()	SplitSingleQuery, SplitBatchQuery
kbase_sumq()	g_kbase_sumq()	SplitSingleQuery

• Matching renames: parameters, locals and private members that hold these values. This includes the optional k1xsumq pointer argument of BatchQuery and the SymphonyQG search entry points, which is now g_k1xsumq. Argument order and types are unchanged.
• Stable core API v1 (core/api.hpp): unchanged. Only a private member of its prepared query is renamed.
• No other changes: factor values, numerical behavior, Python API, saved-index formats and dependencies stay the same.
• Known downstream impact: zvec's tests/core/algorithm/ivf_rabitq/ivf_rabitq_streamer_test.cc calls kbxsumq(). It needs a one-line change to g_kbxsumq() when zvec updates its RaBitQ submodule. A code search finds no zvec use of k1xsumq or kbase_sumq.
• Release notes: the next release notes should list the rename table above.

The second commit fixes documentation naming found in a review of identifiers in the docs:

• estimator.md: used F_add, F_rescale, F_error, G_kBxSumq, … which do not exist in the code. It now uses the public names: f_add/f_rescale/f_error/f_add_ex/f_rescale_ex and g_add/g_error/g_kbxsumq/g_k1xsumq.
• Example variables: estimator.md and quantizer.md now use estimated_distance, error_bound, lower_bound and upper_bound, with _ex for the full-code refinement. They previously mixed lb_dist, est_dist_ex, low_dist_ex and ex_*.
• Cluster-ID arrays: the IVF page and the IVF/HNSW indexing samples call them cluster_ids instead of gt_type/cids. The unused gt_type alias in sample/cpp/symqg_indexing.cpp is removed.
• Python IVF example: passes max_elements, the Python parameter name, instead of num_points.
• docs/development.md: the naming table drops node_id, which no code uses.

The third commit fixes a flaky test that failed in the first ASan/UBSan run of this PR. HnswConstructionTest.ParallelConstructionProducesSearchableIndex required each of the first 16 points to find itself after an 8-thread build. HNSW does not guarantee connectivity: in 500 local 8-thread builds of the same data, about half left 1–2 of the 512 points unreachable, while 1- and 2-thread builds left none. The test now queries every point, checks that each result is a valid ID with a finite distance, and requires at least 99% self-recall (at most 5 misses; the local worst case was 2). That unreachable points appear in about half of the parallel builds may itself be worth tracking separately.

Testing

• Release build with tests, samples and Python bindings, native optimization OFF (GCC 14): ctest --no-tests=error 441 passed, 1 conditional skip (ThreadCount.BoundParallelForUsesRuntimePlacementAndPropagatesErrors).
  • The build emits 14 GCC warnings (-Warray-bounds, -Wunused-result), all on lines this change does not touch. I did not build an untouched baseline to compare against.
• ./scripts/check-tidy.sh (clang-tidy 23, full configured set): pass.
• ./scripts/check-isa-symbols.sh on the portable build: pass, 42 translation units.
• ./scripts/check-format.sh, git diff --check, and python -m mkdocs build --strict --config-file docs/mkdocs.yml: pass.
• Updated HNSW test: 500 repeats each with 8 and 16 threads (Release), plus 300 repeats under ASan/UBSan restricted to 2 CPUs: no failures. Full suite after the change: 441 passed.
• Not run: Python tests. No binding source or Python-visible behavior changed.