[CMGS]

After a sandbox moves, a direct find_iter() call still targets the old owner on every retry. Unlike find(), the iterator never called _rebind, leaving the handle stale after an owner 404 or refused connection. This closes the P2 found after #483.

find_iter() now opens its stream through the existing _open_stream, which follows one move: on an owner 404 or a refused connection it looks the sandbox up and retries once at the new owner, as the Go FindSeq does through retryMoved. A moved owner shows only at the dial or the upgrade reply, both before the first frame, so the retry replays no match. find() collects find_iter() and drops its own _follows_move, so a call still makes one lookup and one retry at most. _open_stream with an empty expect returns the first frame of any type.

The regression uses in-process fake nodes and the existing relay fixture, with no sleeps or timing assertions. It covers both owner 404 and connection refusal, for find_iter() and find(): the first call returns the matches from the new owner, with one lookup and one upgrade there.

Validation on Python 3.12:

• Negative control with the previous find_iter: both find_iter cases fail; both find cases pass.
• Focused migration and callback-error checks: 6 passed.
• Full Python SDK suite: 289 passed.
• Ruff and mypy at the CI pins (ruff 0.15.22, mypy 1.20.2): passed.
• Repository-wide ruff format --check . and ruff check .: passed.
• SDK strict python -m mypy: passed, 11 source files.
• git diff --check: passed.

LOC: production +7/-6 (net +1); tests +22. No dependencies or public API changes. No hardware migration run; this validates the SDK routing and retry boundary.