[BeautyyuYanli]

This PR is made with gpt-6.1-sol (medium) (codex), while I'm responsible for all the changes.

Fixes ENG-1226 — E2E Agent startup fails without a local sandbox auth token.

External Runtime E2E currently exits before Cucumber because Agent's required local sandbox token defaults to an empty string. Resolve one credential per managed run: preserve an explicit E2E token or Agent credential, otherwise generate 32 cryptographically random bytes for the self-managed sandbox. Forward the same token through the shellctl and Agent setup children and Cucumber, including Agent's canonical and legacy environment aliases. Docker receives the token through its environment rather than command arguments.

Explicit sandbox endpoints use their existing service and require its real credential. The runner does not start or stop that sandbox. Production validation remains unchanged, and self-managed CI needs no new GitHub Secret.

Validation:

• All 68 E2E owner tests pass, including 19 focused environment/runner tests. vp check e2e, E2E TypeScript checking, git diff --check, and staged commit checks pass.
• Real source-built shellctl and Agent start with no configured token and with an explicit token. Verified matching credentials in actual child processes, Agent OpenAPI 200, authenticated shellctl requests 200, missing/wrong credentials 401, and no token in startup logs.
• A targeted External Runtime Cucumber scenario passes initialization, launches Chromium, creates/logs in the admin, and executes scenario steps. It exits 1 at the unavailable E2E Agent Decision Chat Model fixture (openai/gpt-5.5, llm): printed result is 1 failed scenario, 5 passed / 1 failed / 10 skipped steps. Full model-generation coverage therefore remains unverified. Local validation reuses a production Web artifact built with native TypeScript stripping disabled to work around this host's Next config import-resolution issue. All verification-owned services are cleaned up.

From Codex