error_info drops the exception type — KeyError('price') surfaces to agents as bare "'price'"

## Summary

The error surfaced to an agent for a failed action carries only `str(exception)`. For `KeyError('price')` that is the bare string `"'price'"` — the exception **class is gone**. An agent picking a fix from the error cannot distinguish `KeyError` from any other exception that stringifies the same way (and in our case `KeyError` vs `TypeError` map to two *different* patches).

## Repro / evidence

From the same `agent-mediated-fork` example. Locally (in-process) the failure is readable:

```
[iteration 1] workflow failed with: KeyError: 'price'
```

But on a remote run, read back from the run's failed action through the SDK (`details.error_info`), the exception type is lost:

```
[iteration 1] toy_sales_pipeline.clean_records failed with: 'price'
```

The agent had to fall back to matching on the *failed task's name* to choose the right patch, because the error text alone was ambiguous.

## Proposed fix

Include the exception class in the propagated error info — e.g. a dedicated `error_type` field, or format the message as `f"{type(e).__name__}: {e}"` at capture time — so `details.error_info` for `KeyError('price')` contains `KeyError`.

## Acceptance criteria

A task raising `KeyError('price')` produces error info in which a downstream agent can read the type `KeyError` without inferring it from the message.