Harbor: preserve verifier evidence and scoring when Kernel cleanup fails

Reproducible control-flow defect

harbor_adapter.test_script("kernel") emits set -euo pipefail and calls kernel-browser.py finalize before copying /data to /logs/verifier/data, running verify.py, or cleaning up email.

In kernel-browser.py:cmd_finalize, a BrowserRuntimeError from provider.cleanup(...) returns 1. Consequently a provider deletion outage aborts the verifier script before scoring or copying verifier evidence. The failure branch writes state/private metadata but skips _write_lifecycle, so the normal /data lifecycle artifact is also absent. This is established by code inspection, not a live provider test.

Acceptance criteria

• Separate scoring, replay collection, browser deletion, and email cleanup outcomes; guarantee cleanup attempts and preserve diagnostic artifacts even if another phase fails.
• Persist a public lifecycle failure artifact before returning on every provider-error path.
• Define how Harbor represents cleanup infrastructure failure without silently converting it into agent failure or hiding it behind an ordinary success reward.
• Fault-inject replay, deletion, verifier, and artifact-copy failures; assert evidence and final exit/result semantics.
• Cover failure after successful setup but before verification, where the setup EXIT trap has already been removed; document provider TTL fallback if cancellation cannot invoke cleanup.

Child of #331. This is a lifecycle/result-contract fix, separate from #346's readiness work. It does not reopen deferred CI #350 or require publishing an image.

@Perry2004 please review the failure semantics and cleanup ownership before implementation.

Start with harbor_adapter.test_script("kernel") and kernel-browser.py:cmd_finalize, then trace the calls to verify.py, /data, and /logs/verifier/data. Map the separate scoring, replay, deletion, email, and artifact-copy outcomes before deciding the lifecycle failure contract. Done means fault-injection coverage for each failure path, preserved public evidence, and documented TTL fallback behavior.