fix: IVF nearest-partition serialization test returns a different class ID

[vsag-bot]

Summary

Corrected the flaky IVF nearest-partition tests to compare equivalent approximate searches and validate classification preservation across serialization.

Root Cause

The serialization test compared ClassifyDatas using ef_search=10 with a restored HGraph queried using ef_search=20. Approximate searches with different breadths can legitimately return different partition IDs, with numerical and platform differences affecting how often this appears; serialization was not shown to be defective.

Changes

• Compare ClassifyDatas results directly before and after serialization using identical inputs and settings.
• Align the basic direct-router check with ClassifyDatas' ef_search=10 configuration.
• Kept .vsag-ai-fix-progress.json valid and untracked.

Validation

• make asan COMPILE_JOBS=4 EXTRA_DEFINED="-DDOWNLOAD_DIR=$VSAG_THIRDPARTY_DOWNLOAD_DIR": passed - Built the complete ASan/UBSan unit-test binary using the local dependency cache.
• ASAN_OPTIONS=detect_leaks=0 ./build/tests/unittests "[IVFNearestPartition]" -r compact: passed - Both focused tests passed with 2,004 assertions.
• 50 repeated post-fix serialization test runs: passed - All independently trained and serialized cases passed.
• git diff --check && clang-format-15 --dry-run --Werror src/algorithm/ivf/ivf_nearest_partition_test.cpp: passed - Diff and formatting checks passed with clang-format 15.0.7.

Residual Risks

• The exact x86/MKL CI image was unavailable locally.
• LeakSanitizer was disabled during runtime tests because the local ptrace environment prevents its process inspection; AddressSanitizer and UBSan instrumentation remained enabled.

Base branch: main

Fixes #2561

Prepared by vsag-bot after automatic effort ordering and AI repairability preflight.

[vsag-bot]

/label status/waiting-for-review
/waiting-on reviewer
/request-review @jiaweizone
/request-review @wxyucs
/request-review @inabao

[vsag-bot]

AI PR state: status/ai-working
Current owner: @vsag-bot
Reason: checks concluded failure
Head: 43ea063

Automatic merging is disabled. A maintainer performs the final merge.

[mergify]

Merge Protections

🟢 All 3 merge protections satisfied — ready to merge.

Show 3 satisfied protections

🟢 Require kind label

• label~=^kind/

🟢 Require version label

• label~=^version/

🟢 Require linked issue for feature/bug PRs

• body~=(?im)(?:^|[\s\-\*])(?:close[sd]?|fix(?:e[sd])?|resolve[sd]?)\s*:?\s+(?:#\d+|[\w.\-]+/[\w.\-]+#\d+|https?://github\.com/[\w.\-]+/[\w.\-]+/issues/\d+)

[LHT129]

The fix correctly addresses the root cause: the serialization test was comparing ClassifyDatas results (which internally uses ef_search=10) against KnnSearch with ef_search=20. Different search breadths in approximate HGraph search can legitimately return different partition IDs, making the test flaky rather than indicating a serialization defect.

Changes review:

1. Basic Test (ivf_nearest_partition_test.cpp:53-56): Aligning ef_search from 20 to 10 so the KnnSearch comparison matches the internal ClassifyDatas configuration. The added comment is helpful since the choice of 10 is non-obvious.

2. Serialize Test (ivf_nearest_partition_test.cpp:96-97): Replacing the per-element KnnSearch loop with a direct ClassifyDatas comparison is the correct approach — it tests that serialization preserves classification behavior using identical inputs and settings, which is what the test should verify.

The changes are minimal (4 additions, 18 deletions), focused, and correct. No issues found.