[Epic #7744 2/6][B4-E1] Preserve host-local inference lifecycle authority

[ericksoa]

Outcome

Persist and re-prove immutable provider-owned host-local inference authority through every supported lifecycle transition.

Dependencies

• Parent epic: [Epic] Support native Podman with buildless managed onboarding #7744
• Completed B4-D foundation: feat(runtime): add dormant Podman GPU and local inference (#8061) #9138 and PR feat(runtime): add dormant Podman GPU and local inference #8061
• Delivery PR: feat(runtime): preserve host-local inference lifecycle #9123

Acceptance criteria

• Persist the inference receipt and route before shared-state commit.
• Bind sandbox, agent, runtime provider, engine, endpoint, immutable image, runtime, model, route, lifecycle generation, and NVIDIA CDI authority.
• Make registration and route ownership compare-and-swap safe across retries and reconciliation.
• Rebind and re-prove authority after destroy, snapshot, clone, rebuild, backup, restore, and recovery.
• Use exact immutable identifiers for cleanup.
• Fail closed on missing evidence, authority drift, name reuse, ambiguous ownership, incomplete creation, incomplete rollback, or indeterminate cleanup.
• Cover OpenClaw, Hermes, LangChain Deep Agents Code, Ollama, NVIDIA NIM, and vLLM.

Scope boundary

• Keep Podman absent from production registration and public support claims.
• Keep --temp-managed-runtime hidden, default-off, and undocumented.

Delivery position

2 of 6 remaining epic batches. PR #9123 closes this issue when B4-E1 merges.