Honour skip_module_check when filtering failures by extension module

[dionmcm]

Port of upstream 051e87e (VAL-531).

What upstream fixed

An extension validation discards any failure whose moduleId is outside the configured module list. That is right for assertions reporting failures against arbitrary components, but wrong for assertions that have already worked out which modules their failures concern — their rows were dropped even though the assertion had deliberately selected them. Upstream lets an assertion mark its own rows exempt by writing qa_result.skip_module_check, which the extractor then honours.

Why this matters to us concretely

It is not only a filtering improvement — on the catch-up corpus it is currently a hard execution failure. Two of arm F's five execution errors are exactly the assertions whose INSERT names this column:

eb7ccd3f  Term exists exactly once in each language refset for each concept in delta.
31f5e2c8  There is only one member id per description per dialect in the language refset snapshot file.

The column does not exist, so the insert fails and the assertion cannot run. This PR creates it — Hibernate builds it from the QAResult entity under ddl-auto=create, so schema and reader stay in step.

Why it is safe to land now, ahead of the corpus

Verified against both corpora:

check	result
assertions writing the column in the pinned corpus fad36466	0
assertions writing it in the catch-up corpus 0160dd2	4 INSERTs across 3 files
any of those among the 544 assertions production runs?	no (checked against daily-rvf 15945's report)
INSERTs using positional VALUES that a new column would break	0 — every INSERT in both corpora names its columns

So on today's nightly the column reads back NULL, getBoolean yields false, and every row goes through the module filter exactly as before. Landing the engine change first means the corpus bump does not have to carry an engine change with it.

Note on fidelity to upstream

MysqlFailuresExtractor has diverged here (Collectors.toList rather than .toList(), inlined limit), so the predicate is ported by meaning rather than as a literal diff. The entity and DTO changes match upstream exactly, including making FailureDetail Serializable — which also gives the pre-existing transient on tableName a meaning it did not have.

Testing

mvn test — 88/88 green. The behavioural claim is under test in a full local run (arm H = this branch + catch-up corpus), whose prediction is execution errors 5 → 3, clearing exactly the two assertions above.

🤖 Generated with Claude Code

https://claude.ai/code/session_018aYBXcyiozxQQGhzzhS7KG