Assert what a failure carries rather than inspect it

[tamnd]

The conditions item of the practice scorecard, which reads: every failure carries the standard's condition code, the place, and the doc URL, checked by a test rather than promised in a page.

Two of the three were not checked here, and the third was checked in a way that could not fail.

The assertion that passes when the field is missing

e.position()
    .ifPresent(
        p -> {
          assertTrue(p.line() >= 1, "line " + p.line());
          ...
        });

That is green when there is no position. So was the caret assertion beside it. The one outcome worth failing for is the one it lets through, and a suite with a few of these in it stops being evidence.

Where a field is genuinely optional the case that has it and the case that does not are now two tests, and each one says which it is. A syntax error has a position and the test asserts it. A connection used after close has none, and the test asserts that too, because a misuse that borrowed a condition code would send a reader to a page about something else.

The doc URL

Never checked anywhere. It is now asserted to be the base plus the code exactly as the code is written:

assertEquals(DOCS + e.code().orElseThrow(), e.docUrl().orElseThrow());

Exactly, because a condition class can have a letter in it, 01G11 for one, and a URL that lower cased the code would be a 404 that nobody notices until somebody clicks it.

What else went in

• the offset is where the token starts, so a caller who still has the text can slice at it, checked by slicing
• the caret sits under the column, checked by the index of the ^
• a name nothing bound is 42002 and not merely something in class 42
• a connection used after close carries MISUSE_CLOSED, no code and no page, deliberately
• a database that is not there names the path it looked for, since that is the one thing the class does not tell you
• the condition classes are subclasses of the one, and the one is unchecked
• a diagnostic for a call that never reached the engine leaves every field empty rather than filling it with a zero that reads like a place in the text

Checked

On a server, against libzu built from the engine at HEAD.

• mvn test -pl zudb-ffm -am -Dtest=ErrorTest: 13 of 13
• mvn test -pl zudb-jni -am -Dtest=ErrorTest: 13 of 13, same cases through the other provider
• mvn test over the whole reactor: build success, every module green, both providers

This takes zu-java from 25 to 35 of 90 on practice. The two items it already held and nobody had recorded are reference, since the javadoc is generated and published with the release, and idiom, since the API is records, a sealed Value, AutoCloseable on every handle, Stream<Row> and unchecked exceptions. The scorecard said zero for all ten, which was never an audit, only a row nobody had filled in.