[mstaats-nvidia]

A job receipt is written before case validation. Add a regression test proving that a validation failure does not let sealed jobs bypass validation on resume.

The fake Harbor runner emits a trial with an incorrect task checksum before sealing. The test checks that both attempts fail real validation, all three job receipts exist, resume never invokes Harbor, no results index or experiment manifest is published, and saved files remain byte-identical. Production code is unchanged.

Follow-up to the receipt-ordering review concern on #106.

Validation:

• Full repository suite: 3,012 passed, 29 skipped (unavailable/opt-in Gym, Node/browser, Docker, Ratchet, Windows and live-model checks).
• Execution suite: 21 passed, 1 opt-in Docker skip.
• Deliberately skipping validation on resume makes the new test fail.
• The same test passes on feat(metric-benchmark): separate cases, execution, and analysis configuration #157 at 12c9718 with its load_execution(cases.yml, execution.yml) fixture-loading adjustment; that branch was not modified.
• Ruff/format, type checks for Python 3.12/3.13, copyright, license, and whitespace checks passed.