Expose run metrics for long background backfills

[chubes4]

Context

Data Machine owns flows, pipelines, jobs, batch scheduling, and processed items. At 10k+ wiki articles and graph backfills, operators need run-level visibility beyond individual job status.

Problem

Large backfills need durable metrics: items processed/skipped/failed, retries, duration, provider/source, flow/pipeline/agent/root context, and cost/token usage where available. Without this, a multi-day pipeline can fail quietly or produce poor-quality output before anyone notices.

Proposed scope

• Add generic run metrics emitted by flow/pipeline/system task execution.
• Track counts for processed, skipped, failed, retried, staged actions, and child job totals.
• Track duration and timestamps for start/end/last activity.
• Add optional token/cost fields when AI providers expose them.
• Expose metrics via CLI/REST/abilities.

Acceptance criteria

• A background flow can report progress and failure rates without domain-specific code.
• Metrics work for pipeline batches and system-task fan-out.
• Intelligence can display root/topic/source progress using generic DM run data plus its own brain context.