fix(engine): stop non-LLM steps constructing a provider for a context window

[Jason Robert (jrob5756)]

Fixes the two failing jobs on main: Web BG Smoke (windows-latest) (failing 3 of the 4 runs since it was added in #450) and a flaky Test (Python 3.13, ubuntu-latest).

1. A step with no model was constructing a provider

_execute_loop built every agent_started payload with a context_window_max resolved through _get_context_window_for_agent, which resolves a provider — and ProviderRegistry builds providers lazily. A wait / set / script / terminate / human_gate step has no model, so that construction bought a guaranteed None.

It is also charged to limits.timeout_seconds, because it runs inside the engine's timed loop. Profiling a provider-free wait workflow, the first step spent 0.434s of the budget building a Copilot SDK client:

gap workflow_started -> first agent_started: 0.451s # before
gap workflow_started -> first agent_started: 0.000s # after

On a cold Windows CI runner that (plus general per-step overhead) consumed enough of examples/wait-smoke.yaml's 3s cap to time the workflow out. The child then exits non-zero, and the launch gate reports it as Background process exited before the workflow started (code 1) — failing the smoke job for something it does not test. In-engine time for that workflow drops from 1.46s to 1.01s, of which 1.00s is the sleeps it is meant to perform.

is_llm_agent was already computed two lines above, so the fix gates the lookup on it. Provider-backed agents are unaffected; the two completion-side call sites already run only in the LLM branches.

TestNonLlmStepsSkipTheProviderLookup pins it with a spy on ProviderRegistry.get_provider — verified to fail on the unfixed engine, so it is not vacuous.

2. test_a_scan_failure_does_not_wedge_the_screen was racing its own poll timer

The two TestRunsScreenGuardRecovery tests asserted _refreshing is False a single instant after settle(), while shrinking POLL_INTERVAL_SECONDS to 0.05s — so a tick landing during settle()'s trailing pilot.pause() legitimately set the flag back to True before the assertion read it (assert True is False). Reproduced locally at ~20%.

They now wait for the flag to clear, which is what the finally actually guarantees: without it the flag latches True for the session and the wait times out, so the regression is still caught. 20/20 clean, from 2/10 failing.

3. Headroom for the smoke fixture

examples/wait-smoke.yaml's 3s cap left no margin on a cold Windows runner even after fix 1, so it is now 15s, with examples/README.md and the ci.yml job comment updated. The timeout path the example demonstrates is unchanged — drive it with a larger --input middle_duration_ms.

Validation

• make check (ruff, ruff format, ty) — clean
• tests/test_engine + tests/test_fleet — 1602 passed, 1 skipped
• make validate-examples — exit 0
• Real conductor run examples/wait-smoke.yaml --web-bg — launches, reaches workflow_completed in 1.013s in-engine, conductor stop cleans up