fix: complete span telemetry on shim and exhaustion terminals

The prompt-shim early return and the maxSteps-exhausted return now attach the same llm.usage.* and gen_ai.* root-span attribute pairs as the other generateText terminals; streamText's teardown block already covers its own paths. The usage-only stream sentinel reads service_tier through the typed StreamChunkExtras surface instead of inline casts. Human-node judge failures keep the fail-safe fall-through to the human interrupt and now warn with the underlying error so judge config mistakes stay visible. Refresh the agency example model and correct the gated-job comment.

Author: jddunn