Add coverage for successful generalization gap calculation

[harrythomasben]

Problem

The runner has a successful path that computes generalization_gap only when both tuned and held-out partitions score, but there is no direct test pinning that calculation.

Proposed coverage

Add a tests-only regression case with mocked successful partitions and assert the reported gap equals tuned minus held-out composite means.

Found by locally pulling upstream/test and reviewing the generalization runner path; this is not based on an existing issue.

[vanguarstew]

This fits the roadmap as a direct prerequisite for the current TEE benchmark scoring work: pinning the successful generalization_gap calculation ensures the deterministic scoring layer is reliable before we compose it with Polaris. The proposed test is tests-only and low-risk, so it's in scope. Please add a regression test that mocks successful tuned and held-out partitions and asserts the gap equals tuned minus held-out composite means, and ensure it runs under VANGUARSTEW_OFFLINE=1. Next step: open a PR referencing this issue with the test and a passing CI run.