[BUG] #680 idle detector flags a session idle while a foreground subagent works

[tronghieu]

Description

The #680 idle detector keys on the main session transcript only (_transcript_activity_key, src/bmad_loop/adapters/generic.py:1750, unchanged on main @ d09a324a6). Claude Code writes a subagent's turns to its own file, ~/.claude/projects/<project>/<session-id>/subagents/agent-*.jsonl. While a foreground subagent works, the parent transcript does not change, so the detector journals session-idle for a session that is busy.

Steps to reproduce

1. Claude adapter, dev_stall_grace_s = 600.
2. Run a story whose dev session hands a slice to a foreground subagent that works for more than 600 s.
3. Read the run journal.

Expected behavior

No session-idle while any subagent transcript of the session is growing.

Actual behavior

From a real run (bmad-loop 0.13.1):

{"kind": "session-idle",   "task_id": "...-dev-1", "idle_s": 604.737, "threshold_s": 600.0, ...}
{"kind": "session-active", "task_id": "...-dev-1", "idle_s": 846.486, ...}

The session was running a foreground subagent for those 846 s. The notice is observability only, but it reads as a stuck session in the journal and the TUI.

Suggested fix

Fold the newest (mtime_ns, size) of <transcript dir>/<session-id>/subagents/*.jsonl into the activity key, stat-only, so usage_parser = "none" profiles keep working.

Environment

bmad-loop 0.13.1 (68645cb), Claude Code adapter, macOS.