test: cover eval_set bundling after retries

[deepujain]

This PR contains:

• New features
• Changes to dev-tools e.g. CI config / github tooling
• Docs
• Bug fixes
• Code refactor
• Tests

What is the current behavior? (You can also link to an open issue here)

Issue #2795 reports that eval_set(..., bundle_dir=...) can attempt to create the output bundle on each retry. With the default bundle_overwrite=False, a duplicate bundle creation would fail once the first retry attempt has already created the output directory.

The current test coverage only verifies bundling when all logs are already complete; it does not exercise a retrying eval set with a real bundle output directory.

What is the new behavior?

• Adds a focused regression test for eval_set() batch retries with bundle_dir.
• The test makes the first eval-set attempt fail, retries successfully, and writes a real bundle with bundle_overwrite=False.
• The assertions verify the bundle exists after completion and contains only the successful eval log, so a duplicate bundle creation during retry would fail the test.

Does this PR introduce a breaking change? (What changes might users need to make in their application due to this PR?)

No. This is test-only coverage for the existing eval-set bundling behavior.

Other information:

Validation:

• Targeted:
  • uv run pytest tests/test_eval_set.py::test_eval_set_bundle_after_retries_once -v
    • Passed. Exercises the reported retry path directly: first batch attempt fails, retry succeeds, and the bundle is created once with overwrite disabled.
  • uv run pytest tests/test_eval_set.py::test_eval_set_bundle_when_all_complete tests/test_eval_set.py::test_eval_set_bundle_after_retries_once -v
    • Passed. Confirms the existing all-complete bundle behavior and the new retry regression coverage both hold together.
• Regression:
  • uv run make check
    • Passed. Runs ruff, formatting checks, and mypy across source and tests.
  • uv run make test
    • Passed. Full pytest suite completed successfully with 7,921 collected tests.

CI/CD coverage expected:

• The standard Build workflow should run ruff, mypy, package inspection, and pytest on Python 3.10 and 3.11.
• The Build Log Viewer workflow should still run schema/type and dist validation; this PR does not touch viewer assets.
• No sandbox tool files or version gates are affected.

Closes #2795.