fix: wait for graceful CLI exit on close to persist session transcript

[n0isy]

Summary

• close() now waits up to 5 seconds for the CLI process to exit gracefully after closing stdin, instead of immediately sending SIGTERM
• Falls back to SIGTERM only if the process doesn't exit within the timeout
• This ensures the CLI's async write queue has time to flush the session transcript to disk

Root Cause

The CLI uses enqueueWrite() + setTimeout() to batch-write session transcript entries to the JSONL file. When close() sends SIGTERM immediately after closing stdin, the write queue hasn't drained yet. The session file ends up containing only a queue-operation entry — no user/assistant messages. When --resume is later used with that session ID, the CLI can't find the conversation history and silently creates a new session instead.

How the Fix Works

The CLI already detects stdin EOF and initiates a clean shutdown, flushing its write queue before exiting (typically under 1 second). This change simply gives it time to do so:

# Before: immediate SIGTERM after closing stdin
self._process.terminate()

# After: wait for graceful exit, SIGTERM only as fallback
try:
    with anyio.fail_after(5):
        await self._process.wait()
except TimeoutError:
    self._process.terminate()

Verification

Tested with a script that:

1. Creates a session via ClaudeSDKClient, sends a query, disconnects
2. Resumes the session with resume=session_id, sends a follow-up query

Before fix: Resume fails silently — new session ID returned, error_during_execution, 0 turns.
After fix: Resume works — same session ID, conversation context preserved.

Also verified via direct subprocess test that the CLI exits naturally in ~0.5s after stdin EOF with the transcript fully flushed to disk.

Fixes #555

🤖 Generated with Claude Code