Add headless sandbox smoke coverage and verified teardown

[Trecek]

Add a headless sandbox smoke test that exercises fleet orchestration and one smoke-task worker through a canary edit, configured sandbox test gate, branch push, and a PR closed without merging. The harness stages the recipe, assigns an exact per-run branch before launch, and verifies the PR's complete diff, immutable head/ref agreement, and executable canary.

The workflow removes and verifies the exact model container before starting independent remote cleanup, including failures that push a branch without creating a PR. Recipe expected failures match exact structured envelope identities; cleanup errors remain fatal. The PR also adds catalog triggers, reproduction instructions, lifecycle/fault regressions, and production fleet ingredient-binding coverage.

base_branch follows fleet's existing config authority using the sandbox project's configuration. This is a necessary adjustment to the plan: fleet strips caller overrides before required-ingredient materialization.

Validation:

• Configured test_check: 16,284 passed, 207 skipped, 12 xfailed.
• pre-commit run --all-files: passed.
• Real container descendant timeout check: passed; delayed writer stopped before cleanup, exact container/CID absence verified, incoming failure preserved.
• Live MiniMax CI: passed with the recorded expected_failure for Fleet L3 recipe bootstrap fails with a retired kitchen generation #5250 and verified post-match cleanup. The assigned branch is absent, no open PR remains, and removal/cleanup exit codes are zero.

A clean worker/canary/PR live run remains pending the retired-generation kitchen bootstrap defect in #5250. Its observed exact signature is linked in the catalog and must be removed after the defect is fixed. The separate failure-reporter defect found during validation is tracked in #5249.

Refs #5232.