GLM: Fix linear constraints not enforced and NPE when combining beta + linear constraints

[Copilot]

GLM's constrained solver (fitIRLSMCS) silently ignored linear constraints due to premature early exit, and threw an unhandled NullPointerException when beta_constraints and linear_constraints were combined.

Investigation Findings

No code changes were ultimately committed. This PR documents the root-cause analysis from the investigation session.

Bug 1 — Linear constraints not enforced

Root cause (GLM.java, fitIRLSMCS() ~line 2293): The augmented Lagrangian outer loop exits on the very first iteration via the standard convergence check:

// PROBLEM: when lambda=0, ck=0.01 (initial), augmented Lagrangian ≈ unconstrained objective.
// Solver finds unconstrained solution quickly → _betaDiff tiny → converged()=true → progress()=false → early exit.
// Lambda and ck are NEVER updated, so constraints are never enforced.
if ((!progress(betaCnd, gradientInfo) && !gradSmallEnough) || done) {
    checkKKTConditions(betaCnd, gradientInfo, iterCnt);
    if (_betaConstraintsOn) bc.applyAllBounds(_state.beta());
    return;
}

Required fix: Strip the (!progress(...) && !gradSmallEnough) short-circuit from the early-exit condition — only exit on done. The outer loop must be allowed to update λ and c_k before convergence is evaluated:

done = stop_requested() || (_state._iter >= _parms._max_iterations) || _earlyStop;
if (done) {
    checkKKTConditions(betaCnd, gradientInfo, iterCnt);
    if (_betaConstraintsOn) bc.applyAllBounds(_state.beta());
    return;
}
progress(betaCnd, gradientInfo); // update state only, don't use as exit condition

Bug 2 — NPE when combining beta_constraints + linear_constraints

Exact call site not pinpointed, but two confirmed contributing defects in ConstrainedGLMUtils.java:

• countNumConst (~line 850): Integer division state._lessThanEqualToConstraintsBeta.length / 2 returns 0 for a single-sided constraint (length=1) instead of 1.
• printConstraintSummary (~line 378): References state._lessThanEqualToConstraints (the combined field, only populated mid-solver) instead of state._lessThanEqualToConstraintsLinear — potential NPE if called before solver populates the combined field, and causes duplicate constraint display.

Files Requiring Changes

• h2o-algos/src/main/java/hex/glm/GLM.java — Bug 1 fix in fitIRLSMCS()
• h2o-algos/src/main/java/hex/glm/ConstrainedGLMUtils.java — Bug 2 fixes in countNumConst and printConstraintSummary

Two new Python regression tests are already present in the repo:

• pyunit_GH_16312_contrained_GLM_test_large.py — equality constraint enforcement (x2+x3=0)
• pyunit_GH_16312_contrained_GLM_beta_constraint_NPE_large.py — beta + linear combined (lower-bound and upper-bound cases)

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud