CLI `run` command has no way to capture a snapshot of the running preview

[obj-p]

Problem

The MCP server exposes preview_snapshot to capture a screenshot of a running session, but the CLI run command has no equivalent. The snapshot CLI command is one-shot — it creates its own session, captures, and tears down. There's no way to capture the window from a run session programmatically.

This gap makes it impossible to:

• Script before/after comparisons across hot-reload cycles
• Verify visual state in CI without the MCP server
• Debug hot-reload issues (e.g., validating Setup plugin state lost after hot-reload (static linking gives each dylib its own statics) #86 fix) from the CLI alone

Desired behavior

snapshot should reuse an existing running session when one exists for the same file. If run already has a preview window open, snapshot should capture that window rather than booting a new session. It should only create and tear down its own session when no existing session is running.

Possible approaches

• Session discovery: run registers its session (e.g., via a pidfile or Unix socket), snapshot checks for it before creating a new one
• --snapshot-dir flag on run: dump a PNG to a directory after each reload (automatic, no separate command needed)
• Unix socket: run listens on a socket, snapshot (or a new capture subcommand) connects and requests a screenshot

Context

Discovered while testing #91 — the setup dylib hot-reload fix couldn't be visually verified via CLI because run has no capture mechanism.