[hutiefang76]

CanberraDistance can return NaN for opposite finite extremes and zero for same-sign large values because intermediate arithmetic overflows. This fixes MATH-1678 by halving the coordinates only when their finite magnitudes overflow the denominator. Ordinary values and the existing NaN/infinity behavior are preserved. Regressions cover both signs, vectors, near-equal large values against an independent BigDecimal reference, subnormals and signed zero.

Validation:

• mvn -B clean verify: all 14 reactor modules completed successfully; archived Surefire results contain 3,797 tests, with 33 existing skips and no final failures/errors.
• The first full run failed in the existing randomized SimplexOptimizerTest. That class also failed on an unmodified baseline. One complete rerun passed using the project's unchanged retry configuration; no tests or assertions were disabled.
• A clean Java 8 build and the affected CanberraDistanceTest passed all 8 tests. The tests-only baseline reproduced 4 failures before the fix.

AI assistance: DeepSeek and Codex helped implement, review and validate this patch. This description was AI-generated.

[hutiefang76]

The three workflows currently show action_required with no jobs. Could a maintainer approve the CI runs when convenient?

[codecov-commenter]

Codecov Report

❌ Patch coverage is 83.33333% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 87.19%. Comparing base (f554608) to head (7eaa1d0).
⚠️ Report is 384 commits behind head on master.

Files with missing lines	Patch %	Lines
...ons/math4/legacy/ml/distance/CanberraDistance.java	83.33%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@             Coverage Diff              @@
##             master     #336      +/-   ##
============================================
+ Coverage     86.54%   87.19%   +0.64%     
+ Complexity     9787       91    -9696     
============================================
  Files           532      499      -33     
  Lines         35516    33512    -2004     
  Branches       6194     5846     -348     
============================================
- Hits          30738    29221    -1517     
+ Misses         3518     3168     -350     
+ Partials       1260     1123     -137     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[hutiefang76]

The Ubuntu/JDK 17 failure is in CMAESOptimizerTest.testDiagonalRosen; testCircleFitting2 also needed a retry. Both tests and their optimizer implementations are unchanged by this PR.

I ran those same methods on pristine base a0df91d8 with Java 17 on macOS arm64. Both assertions failed there too and recovered under the existing retry configuration (2 methods, 2 flakes). This is separate baseline evidence, not a reproduction of the Ubuntu runner or a green hosted build. I have left the tests, tolerances, seeds and CI configuration unchanged.

_{AI-generated reply. The baseline results were checked against the actual execution log.}

[aherbert]

If you rebase on master it should collect the changes to the github macOS runner.

[aherbert]

I don't think numerator can ever be non-zero if the denominator is zero. This check could be:

denom == 0.0 ? 0.0 : num / denom;

It is then testing for a[i] == b[i] == 0.

[hutiefang76]

Changed to denom == 0.0 ? 0.0 : num / denom in 7eaa1d03. A zero sum of the absolute values means both coordinates are zero, so the numerator check was redundant. The 8 Canberra tests pass, including signed zero and subnormal cases. mvn clean verify also passes locally on Java 17/macOS arm64 (3798 tests, 33 skips).

The branch already includes current master a9aabf60 through merge dadaf4cd, including the macos-latest and Zulu/Java 8 runner changes (56a8f696 and a9aabf60). Both are ancestors of this commit. Hosted CI still needs to run on the new head.

_{AI-generated reply; validation figures are from the local execution logs.}