[leerho]

Adds union-result cross-language (CL) files to the HLL cross-language test. Test-only; no library code changes.

Why

The existing HLL CL files are update sketches only, serialized compact, at a single lgK. That misses two classes of bugs, and both turned up in September 2026:

• The deferred KxQ rebuild can only be reached through a union, and the generator creates no unions.
• Flags bit 32 collided between implementations (REBUILD_CURMIN_NUM_KXQ in Java, FULL_SIZE in C++). It lives in the flags byte, which nothing inspected, and only the compact form was ever written.

What it adds

• generateUnionBinariesForCompatibilityTesting writes six files, hll{4,6,8}_union_java.sk (compact) and hll{4,6,8}_unionupd_java.sk (updatable). They come from a union whose inputs span all three curModes, so coupon updates into a gadget with a pending KxQ rebuild are exercised. One lgK is enough, because neither defect class depends on lgK. The files go to the gitignored generated-files directory, so the repository doesn't grow.
• checkUnionJava, checkUnionCpp, checkUnionGo and checkUnionRust share one check body. For each file it asserts that:
  • flags bit 32 is never set;
  • the compact flag matches the form that was written (HLL_6 and HLL_8 used to omit it);
  • the estimate matches the true count, and is the same for every target type, since an incorrect KxQ rebuild shifts it;
  • curMin and numAtCurMin, which the KxQ rebuild also updates, match Java's own result for the same target type and form (added in review).

No C++, Go or Rust generator writes these files yet. Until one does, those checks warn about the missing files and skip them. Matching generators in the other repos would be follow-ups.

Testing

• -Pcheck_java_files: generates the files and reads them back, 4/4 HllSketchCrossLanguageTest tests pass.
• When it was written, the check was verified by corrupting a CL file: setting bit 32, clearing the compact flag, or changing numAtCurMin, each fails the check.
• mvn verify passes (2320 tests). The default test run is unaffected, because these groups are excluded by default.

Supersedes the never-opened branch hll-xlang-fixtures. This is the same commit, rebased onto main after #771.

🤖 Generated with Claude Code

[proost]

For more robust compatibility test, how about adding comparing "curMin", "numAtCurMin"?

[leerho]

Agreed, those are exactly the fields the deferred rebuild updates, so checking them is more direct than relying on the estimate. Added (b466491): each union CL file's curMin and numAtCurMin are compared with the values in Java's own result for the same target type and form. Only the Java CL files exist today; the C++/Go/Rust checks will start comparing once those languages write union CL files.

[proost]

Thank you!