feat: add JUnit XML exporter and --fail-under CI quality gate

[brainsparker]

What this is

A rebase of #41 onto post-consolidation main. The original conflicted with #40 in cli.py; this branch carries the same change set resolved against the merged consolidation.

What it does

CI-native quality reporting: a junit output format (one testsuite per model, one testcase per golden-set entry; model API error maps to error, unjudged case maps to skipped so judging-disabled runs never false-pass, below-threshold judge score maps to failure) and a --fail-under CLI gate that exits with code 2 when any model's average judge score falls below the threshold. A model with no scores also fails the gate rather than silently passing. Stdlib xml.etree only, no new dependencies.

Validation

• Full suite on this branch: 54 passed, 0 failed (36 post-consolidation plus 18 new JUnit exporter tests)

Attribution

Content originally authored in #41; that PR will be closed with a pointer here.