[insaf021]

Observed on master with the default formats:

new RealMatrixFormat().parse("{{1,2},{}}") -> Array2DRowRealMatrix{{1.0,2.0}}
new RealMatrixFormat().parse("{{1},{2}")   -> null
new RealVectorFormat().parse("{1;2")       -> null

Two separate paths let malformed text through.

parse(String) is documented to throw MathParseException, but both classes detect failure with parsePosition.getIndex() == 0. The parseFixedstring failure paths in parse(String, ParsePosition) leave the index past whatever was consumed, so the caller gets null instead of the exception. ComplexFormat already tests result == null; the same test is now used here.

In RealMatrixFormat, a row that has its prefix but no component is treated as the end of the matrix, and the unconsumed } is then matched as the matrix suffix. The empty row disappears and the result has the wrong dimensions. Once a row has been opened after earlier rows it now needs a component, so it fails like any other invalid component. Formats with an empty row prefix have no opened row to reject and still accept a closing row separator; a test pins that.

Regression tests in RealMatrixFormatAbstractTest and RealVectorFormatAbstractTest fail without the patch. mvn clean verify on commons-math-legacy and its dependencies passes on JDK 17 (3278 tests in legacy), as does checkstyle:check. This is separate from #338, which fixed the CompositeFormat cursor helpers.