fix(execution-core): Make failed runs close as Failed in Temporal

[dawidaksamski]

Problem

WB-428: when a node failed under the fail error policy, runGraph wrote the failure to
our DB and then returned instead of raising. Temporal therefore closed the workflow as
Completed — so the product said failed while the engine said completed about the same
execution, and anyone reading history in Temporal saw a success.

Fix

runGraph now returns a RunGraphOutcome ({ status: 'completed' } | { status: 'failed', error })
instead of void, and the Temporal adapter raises ApplicationFailure.nonRetryable when the
outcome is failed.

The raise sits in the adapter rather than the runner for two reasons: execution-core is
deliberately engine-agnostic, and only a TemporalFailure actually fails a Workflow Execution —
a plain Error fails the workflow task and retries it forever. Raising from the runner would
have meant importing @temporalio/* into core.

Non-retryable: a graph failure is deterministic, so replaying it would re-run node activities (LLM spend) and re-emit events. This is the first such declaration in the repo. It does not affect manual reset/rewind, which is independent
of retry policy.