plugin: follow each harness's own transcript layout once the readers land #84

Once #258 and the token-burn stack (#230, #257) have both landed, hook.py's transcript(), seed_cursor() and leave_transcript() need to follow each harness's layout. Today they assume Claude Code's.

• Codex: a subagent's hook carries the parent's transcript_path and an agent_id, and transcript() builds a Claude-style <session>/subagents/agent-<id>.jsonl path that does not exist. The subagent's rollout is its own file, named on SubagentStop as agent_transcript_path, and post already finds it from CODEX_THREAD_ID.
• Gemini CLI: a subagent's chat is chats/<parent session>/<session>.jsonl. Its SessionStart should seed the cursor as Claude Code's does.
• Cursor and Copilot CLI: the hint is left after the tool runs, so a context's first post has none, and the rest name the right transcript only if transcript_path is in the payload.

An integration test should run a session through the hook and then post, for each harness with a reader.