Pitch

We already write pre-turn checkpoints to `~/.deepseek/sessions/checkpoints/latest.json` (the atomic-write work in #355). When `deepseek` starts and a checkpoint exists from a session that didn't shut down cleanly, prompt:

The previous session ended unexpectedly mid-turn. Restore?
[r]estore [s]tart fresh [d]elete checkpoint

Codex parity but better — we already have the checkpoint, we just don't surface it.

Implementation sketch

• `session_manager::load_checkpoint` already exists from quality: schema migration up-path for ~/.deepseek/ records — currently reject-only, no upgrade #350 era.
• On startup, if checkpoint is present AND mtime is within 24 h, show the prompt before the empty composer renders.
• `r` calls into the engine's restore path; `s` clears the checkpoint and proceeds; `d` deletes without prompting again.
• Honour `--no-restore-prompt` CLI flag for users who don't want the dialog.

Acceptance

• Crash-mid-turn → next launch shows the restore prompt.
• Clean exit clears the checkpoint so no false prompt on next run.
• Standard verification gates pass.