Reject unmeasured ML training outcomes

[abrichr]

Summary

• Reject malformed, ambiguous, empty, or non-finite training evidence instead of treating it as success.
• Require measurable optimizer updates and valid checkpoints in the GRPO validation path.
• Add focused regression coverage for false-success cases.
• Make the benchmark dependency and lock file reproducible in hosted CI.

Verification

• 519 passed, 5 skipped in the full non-integration test suite.
• 116 passed in the focused GRPO and false-success suites.
• Ruff lint and format checks passed.
• Wheel and source distribution builds passed.
• The source-availability boundary passed for the built archives.

This change does not request or publish a package release.