chore: bump verifiers for the unbounded model-call timeout

[mikasenghaas]

Bumps deps/verifiers 29e3a0f7 → a298bcfe, picking up two commits:

commit
a298bcfe	fix(v1): restore the unbounded model-call timeout (verifiers#2304)
9ca7f5d6	Fix plain HTTP CONNECT for Docker framework routes (verifiers#2298)

Why

verifiers#2218 consolidated client construction into build_async_openai / DEFAULT_TIMEOUT and, in doing so, replaced the relay client's explicit no-timeout with a 600s read timeout "mirroring the OAI SDK defaults":

-        # No timeout: agentic completions are slow and the rollout timeout is the real backstop.
-            timeout=None,
+DEFAULT_TIMEOUT = httpx.Timeout(connect=5.0, read=600.0, write=600.0, pool=600.0)

That cap reached every model-call path, since DEFAULT_TIMEOUT feeds clients/eval.py as well as build_async_openai, which judge.py and clients/train.py use.

600s is fine for chat-shaped evals and wrong for RL. A single agentic turn on a long SWE trajectory routinely runs past ten minutes; when it does, httpx raises TimeoutException, clients/eval.py maps it to a 504, and the rollout dies as ProviderError: Request timed out — even though the model was still producing. Because it truncates the longest trajectories specifically, the loss is biased rather than uniform.

Observed on the laguna-s SWE RL run on this cluster. verifiers#2304 restores unbounded read/write/pool and keeps connect at 5.0 so an unreachable endpoint still fails fast.

Scope

No lockfile change. verifiers is an editable path dependency (source = { editable = "deps/verifiers" }), and neither commit touches its pyproject.toml, so only the submodule pointer moves. 9ca7f5d6 touches a single file, verifiers/v1/runtimes/docker/egress.py.

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches every model-call path used in eval, judge, and train clients; the timeout change directly affects RL rollout reliability but fixes a biased failure mode rather than adding new logic.

Overview
Advances the deps/verifiers submodule (29e3a0f7 → a298bcfe) so this repo picks up two upstream fixes. There is no uv.lock change—verifiers stays an editable path dependency and only the submodule pointer moves.

The main behavioral change is restoring unbounded read/write/pool timeouts on model HTTP clients (with 5s connect kept for fast failure on dead endpoints). That reverses a regression where a shared 600s cap on all OpenAI-style clients was aborting long agentic turns during RL—showing up as TimeoutException / ProviderError: Request timed out on the longest SWE trajectories even while the model was still streaming.

Also included: a small Docker runtime fix for plain HTTP CONNECT on framework egress routes (verifiers/v1/runtimes/docker/egress.py).

^{Reviewed by Cursor Bugbot for commit 166a0ba. Bugbot is set up for automated code reviews on this repo. Configure here.}