reject mixed stack metadata within reporting repetitions

[anxkhn]

Summary

The llm-d report adapter checks stack metadata only in the first stage file of each repetition. A later stage can name a different backend or component version and still contribute metrics to a comparison labeled with the first stage's backend.

Check the existing canonical stack metadata in every stage file before aggregation. The error identifies the mismatched file. Metric definitions, duplicate-rate warm-up handling and output formats remain unchanged.

Regression fixtures cover backend drift in warm-up and measured stages, non-engine component drift, and drift in a later repetition. A consistent two-repetition fixture checks median output rates, warm-up exclusion, the backend label and Prism bundles. Five regression subcases failed with the original adapter and pass with the fix.

Validation

• Reporting suite: 9 tests passed, including Markdown, PNG, CSV and Prism checks.
• llm-d Python suite: 15 tests passed.
• All five mocked shell harness suites passed.
• ShellCheck and Bash syntax passed for all 15 shell files.
• Python syntax passed for all 18 Python files.
• Standalone repository reference check and git diff --check passed.

These checks ran locally on macOS with isolated Python 3.12.13. Shell tests use mocks and fail-closed tool-presence stubs. No cloud resources, GPUs or live clusters were provisioned.

The existing main-branch Inference tooling run fails during the shell-test step before Python tests. That hosted CI failure is separate from the passing local results above.