SandboxSession kills workers when controller callbacks take >1s (WORKER_CPU_LIMIT hardcoded)

Component

hyperforge.codemode.sandbox (SandboxRunner remote protocol / SandboxSession)

Summary

hyperforge/codemode/sandbox.py hardcodes WORKER_CPU_LIMIT = 1 (line 28). SandboxSession applies it as the wait budget while the controller answers a nested task request, not just while the worker is CPU-bound:

• SandboxSession.run (line 306): await self._start_timeout(WORKER_CPU_LIMIT) — armed while waiting on the worker task.
• SandboxSession._callback (line 361): await self._start_timeout(WORKER_CPU_LIMIT) — armed while waiting for the controller to respond to a SandboxMessage.Request (a task issued by generated code, answered by the SandboxRunner callback).

When the wait exceeds 1 second, _timeout fires and kills the worker process:

async def _timeout(self, timeout: float):
    await asyncio.sleep(timeout)
    self.timed_out = True
    if self.process.is_alive():
        self.process.kill()

Impact

Any legitimate nested capability dispatched through the controller callback that takes more than 1 second (a database query, an HTTP call, a paginated lookup) causes the entire sandbox worker to be killed with RuntimeError("Python agent timeout"). Embedding applications have no way to raise this bound; it is not part of SandboxSettings and not a parameter of run_sandbox_server().

Verified on hyperforge==1.0.0.post310.

Proposal

Move the value into SandboxSettings with the current value as the default (fully backward compatible), e.g.:

class SandboxSettings(BaseSettings):
    ...
    sandbox_callback_wait_seconds: float = 1.0  # replaces WORKER_CPU_LIMIT usage

Worker-runtime enforcement via WorkerExecutionRequest.max_runtime_seconds is separate and should be unaffected.

Acceptance criteria

• The controller-callback wait is configurable via settings/env, default unchanged (1.0s).
• max_runtime_seconds enforcement behaves exactly as before.

Start in hyperforge/codemode/sandbox.py at SandboxSession.run and _callback, then trace how SandboxSettings is exposed through run_sandbox_server. Confirm the callback wait uses a configurable setting while WorkerExecutionRequest.max_runtime_seconds remains unchanged. Done means the default remains 1.0 seconds, settings or environment can raise the callback wait, and the existing timeout behavior is preserved otherwise.