fix(flow): treat agent exceptions as failed steps so retry/fallback engage

[XuQuanxin04]

Summary

• _run_step called _call_agent without catching exceptions. Real agents typically raise on API/network/rate-limit failures (rather than returning a RunResult with error set), and _call_agent re-raises those — directly when no timeout is set, or via future.result() when one is. The exception therefore escaped the retry loop on the first attempt:
  • max_retry never retried a raising agent (the workflow crashed on attempt 1 instead of trying again),
  • fallback_agent was never consulted for a raising agent,
  • with no fallback, flow.run() raised instead of marking the step failed and skipping its dependents.
• Fix: wrap both the primary and fallback agent calls so a raised exception is normalized into a failed step result using the existing [LA-FLOW-...] error convention. The retry loop and fallback path then engage exactly as they do for errors returned in a RunResult, and the flow fails gracefully with downstream steps skipped.

Reproduction (before this patch)

class Flaky:
    def __init__(self): self.calls = 0
    def run(self, query, **kw):
        self.calls += 1
        if self.calls < 3:
            raise RuntimeError("transient API error")
        return "recovered"

flow = LightFlow().step("flaky", agent=Flaky(), max_retry=3)
flow.run("go")   # raised RuntimeError after 1 attempt; never retried

Compatibility

• Existing agent.run("hello") behavior is unchanged for agents that return normally.
• Existing stream=True behavior is unchanged.
• Behavior change is strictly a fix: agents that raise no longer crash the orchestration; they are retried / fall back / mark the step failed, consistent with how RunResult(error=...) failures are already handled. KeyboardInterrupt/SystemExit (not subclasses of Exception) still propagate.

Tests

• python -m compileall -q LightAgent
• PYTHONPATH=. python -m pytest -q tests/test_lightflow.py — 17 passed
• Added three regression tests:
  • test_lightflow_retries_when_agent_raises — a transiently-raising agent is retried and succeeds.
  • test_lightflow_fallback_runs_when_agent_always_raises — a persistently-raising agent falls back.
  • test_lightflow_marks_step_failed_when_agent_always_raises_without_fallback — the step is failed, the dependent step is skipped, and run() returns a failed result instead of raising.

Notes

No API changes. The pre-existing retry/fallback tests only covered errors delivered via RunResult(error=...); this covers the raised-exception path, which is the common failure mode for real LLM integrations.