Add a harness for snapshotting what agents "see" via code-mode (#9342)

The point of code-mode is to give the agent full visibility over what it
just did. There's a lot it can do (run cells, mutate the graph, trigger
reactive cascades, hit state setters) and we didn't have a good way to
capture what an agent observes across those scenarios in marimo-pair.

These changes add a fixture that spawns a real `marimo edit` subprocess
per test and drives it through the full HTTP surface (websocket +
`/api/document/transaction` + `/api/kernel/run` +
`/api/kernel/execute`). Tests snapshot the normalized SSE body, the
bytes the agent's client actually receives.

Seven initial scenarios cover issue #9255: three pass as regression
guards, four are `xfail(strict=True)` pinned to the broken `mo.state`
cascade variants so the fix PR can't land without removing the markers.

Author: manzt