Sub-agent resume and transcript persistence

[bug-ops]

Part of #974

Background

Claude Code assigns each sub-agent invocation a unique ID and stores the full conversation transcript. A subsequent call can resume from an existing transcript, preserving tool results, intermediate reasoning, and multi-turn history. This is essential for long-running background agents that may be interrupted, and for audit/replay.

Current state in Zeph

SubAgentManager tracks status (running, completed, failed) per agent instance via a HashMap<AgentId, AgentHandle>. No transcript is persisted to disk. When an agent completes or the process exits, all context is lost.

Implementation

1. Transcript model: define AgentTranscript { id: AgentId, name: String, started_at: DateTime<Utc>, messages: Vec<Message>, status: AgentStatus }. Serialize to JSON.

2. Storage: persist transcripts under {data_dir}/zeph/agents/transcripts/{id}.json using async write after each turn. Use tokio::fs for non-blocking I/O.

3. AgentId generation: use uuid::Uuid::new_v4() — already a workspace dep.

4. Resume API: add SubAgentManager::resume(id: AgentId) -> Result<AgentHandle> that loads transcript and injects it as initial conversation history into the new LLM session.

5. CLI flag: --resume <agent-id> for the agent run subcommand.

6. Transcript listing: SubAgentManager::list_transcripts() -> Vec<TranscriptMeta> for TUI/CLI display.

7. TTL cleanup: transcripts older than transcript_ttl_days (config, default 30) are deleted on startup.

Acceptance criteria

• Each sub-agent run is assigned a stable AgentId (UUID v4).
• Transcript is written to disk after each turn (not just at completion).
• SubAgentManager::resume(id) successfully restores message history.
• --resume <id> CLI flag works end-to-end.
• Transcripts older than TTL are pruned at startup.
• Unit tests: write transcript, read it back, verify message count.
• cargo nextest run -p zeph-core passes.

Technical notes

• Use serde_json for transcript serialization (already a dep).
• data_dir resolution: dirs::data_dir() + "zeph", configurable via ZEPH_DATA_DIR env or data_dir config key.
• Keep transcript writes non-blocking: spawn a tokio::task::spawn per write, do not await in hot path.
• The Message type is already defined in zeph-llm; reference it from zeph-core via the existing dependency.