SubagentStop hook payload: expose usage tokens and model of the finished subagent

Feature request

The SubagentStop hook receives no .usage.* (input/output/cache tokens) nor model for the subagent that just finished. We instrumented our hook to log these fields on 2026-05-11 and collected 381+ samples: 100% zeros / "unknown" — the payload simply does not carry them (verified across Claude Code 2.1.x).

Why it matters: without per-agent token counts, any cost claim about multi-agent pipelines is unfalsifiable. We run an 11-agent pipeline (plan → implement → qa → security → review) and can only do duration-weighted cost estimates.

Ask: include model and usage (same shape as API responses) in the SubagentStop hook JSON payload — and ideally the resolved subagent_type, which today is also absent (we reconstruct it with a 3-tier resolver, ~64% of our historical entries are unknown).

Environment: Claude Code CLI on macOS (darwin), project-level hooks via settings.json.