GeezerVison

… (#10325)

Adds an object-oriented `Devbox.on_evict(callback)` (sync and async) backed by a
per-client `EvictionMonitor`. The first registration opens a single account-wide
`watch_evictions` SSE stream; incoming notifications are matched against the local
interest set, the devbox is removed before its callback fires (at-most-once), and
the stream is closed once the interest set empties.

Depends on the Stainless-generated `devboxes.watch_evictions` method and
`DevboxEvictionEvent` type, which land downstream once the server endpoint merges.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

- Use the real generated event type DevboxEvictionEventView (the endpoint's view)
  in place of the assumed DevboxEvictionEvent.
- Invoke the callback as callback(devbox, eviction_deadline_ms): the Devbox object
  and the Unix millisecond deadline, instead of passing the raw event.

Requires the main-repo stainless config change that generates watch_evictions as
Stream[DevboxEvictionEventView] (runloopai/runloop#10404).

Co-Authored-By: Claude Opus 4.8 <[email protected]>

The mux SSE route moved from /v1/devboxes/watch_evictions to
/v1/devboxes/evictions/watch (runloopai/runloop#10409) to avoid colliding
with GET /v1/devboxes/{id}. Update the generated request path so the
watch_evictions method reaches the new endpoint. Method surface unchanged.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

Two bugs made devbox.on_evict never fire:

1. The generated watch_evictions() sends Accept: application/json, but the
   endpoint only streams for Accept: text/event-stream — otherwise it returns an
   empty text/plain 200. The monitor now forces the SSE Accept header.
2. The monitor ran the stream once and stopped. The server force-closes on
   leader change / slow consumer (and a long-lived HTTP/2 stream can drop), and
   expects the client to reconnect and re-read the snapshot. The monitor now
   reconnects with backoff until no devbox is still interested.

Adds debug logging around connect / event / reconnect. Verified against dev: a
forced flex drain now delivers on_evict for 20/20 devboxes before suspend.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

Addresses review on #824: AsyncEvictionMonitor still had the pre-fix single-shot
_run, so AsyncDevbox.on_evict would silently never fire. Port the sync monitor's
fixes to sdk/async_eviction.py:
- request the SSE Accept header (text/event-stream) on watch_evictions
- reconnect with capped backoff until no devbox is still interested

Also ruff-format sdk/eviction.py (the sync fix wrapped a call that fits on one line).

Co-Authored-By: Claude Opus 4.8 <[email protected]>