[FEATURE]: Structured usage export for task-level evaluations

Feature hasn't been suggested before

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

I want to compare the same coding task with and without an external retrieval tool. Session totals do not identify the experiment, distinguish missing usage from zero, or establish equivalent task quality.

Would a small structured usage export, reusing the existing session transcript accounting, fit the project? The first version would summarize each provider step into disjoint input, output, reasoning, cache-read, and cache-write buckets, report its accounting scope, and support JSON or CSV. Graph retrieval stays in an external MCP integration.

Related work: #42324 covers multi-step message totals, #45417 covers subagent rollups, #46371 covers hidden calls, and #41915 proposes a human-facing usage report. #47822 proposes returning child-session metrics from the task tool; this proposal instead describes the accounting scope of a complete exported session. #9387 asks for transcript formats; this proposal exports a machine-readable usage summary rather than another transcript representation.

Before making the PR ready, I would like to confirm whether the export belongs under opencode export and how incomplete child/fork usage should be represented.