Rebuild upgrade evals around delivered apps (#99880)
### Why?
Upgrade evals need to verify the app an agent delivers and whether it
communicates security/latest notices correctly.
### How?
- Run six cases with Codex and Claude: same-major and cross-major
upgrades for security and latest, plus one nudge per policy. Exercise
the built candidate against frozen upstream metadata.
- Capture committed, dirty and untracked changes from delivered
worktrees. Require successful candidate invocation and verify manifests,
lockfiles, installed versions, TypeScript, build and HTTP/browser
behavior in a clean Sandbox.
- Require nudge emission and a policy-specific notification or question
about proceeding in the final response. Fail any user-facing
prompt-injection discussion or language discrediting Next.js or its
nudge. Require completed boolean judge verdicts and distinguish failed
checks from invalid execution or grading evidence.
- Reuse the existing agent-eval runner and tools. Patch its lifecycle
for delivery/judge evidence and shared fixture preparation; stop the
preparation VM when its last consumer cancels and delete snapshots that
arrive after cancellation.
- Remove obsolete Future cases and duplicate fixtures/helpers. Generate
starting-app lockfiles during setup and share credential-free tooling
and app preparation.
<!-- NEXT_JS_LLM -->
