feat(v1): progress hook for live traces

[faresobeid]

What

Agent.run(...), Agent.interaction(...) and Rollout(...) take on_progress: Callable[[Trace], None] | None. The interception server invokes it from record_call — the one seam every recorded exchange passes through, after the turn it committed — so the hook sees the live trace each time it gains a ModelCall (and the nodes of that turn). A hook that raises is logged at WARNING and swallowed; a consumer bug never fails a rollout. None (the default) changes nothing. Callers debounce.

Why

on_trace hands out the trace object at mint, but nothing signals when it grows: a live consumer has to poll. In the data-flywheel pipeline (a task factory on verifiers.v1 + the rho harness, ~90 concurrent rollouts), a 2-hour rollout is invisible until it closes — its model calls and turns can only be inspected after the fact, so stuck or runaway rollouts are noticed late.

Validation

• tests/v1/test_trace.py::test_on_progress_observes_live_trace — offline: a loopback chat-completions upstream, an in-process harness that makes 3 model calls through the real interception server, agent.run(task, on_progress=...). Asserts the hook fires once per call with the live (is) trace, len(trace.calls) grows 1, 2, 3, and that a hook raising every time still leaves trace.ok with no errors.
• uv run pytest tests/v1 -m "not e2e": 87 passed. uv run ruff check, uv run pre-commit run --all-files: clean.

Note

Add on_progress live-trace callback to Agent.run, Rollout, and RolloutSession

• Adds an optional on_progress parameter threaded through Agent.run, _run_once, interaction, _EpisodeAgent, Rollout.__init__, and RolloutSession so callers can observe the live Trace after each recorded model call.
• InterceptionServer call-recording handler now invokes RolloutSession.progress after trace content is committed, delivering the current trace to the callback.
• RolloutSession.progress returns immediately when no callback is set and logs ordinary Exception subclasses without re-raising, so a faulty consumer cannot fail the rollout.
• Adds an integration test in test_trace.py using a local loopback provider and ProbeHarness to assert three calls, a shared live trace, callback counts of one through three, and rollout success despite a RuntimeError raised from the consumer.
• Behavioral Change: callbacks that raise non-BaseException errors are swallowed and logged; callers relying on exceptions propagating out of Agent.run will no longer see them.

^{Macroscope summarized e19af4d.}