fix(workflows): fail a fan-in step whose output is not a mapping

[jawwad-ali]

Problem

FanInStep.execute began with:

output_config = config.get("output") or {}
if not isinstance(output_config, dict):
    output_config = {}

Any non-mapping output was silently replaced with {} and the step still returned COMPLETED. Every aggregation key the author declared vanished, and downstream {{ steps.<id>.output.<key> }} resolved to None and interpolated as an empty string.

Note also that x or {} masked the falsy shapes ([], false, 0, '') before the isinstance check even ran.

Reproduction on current main (81bf741)

output=[]       -> status=completed  error=None
output=False    -> status=completed  error=None
output=0        -> status=completed  error=None
output=''       -> status=completed  error=None
output=['a']    -> status=completed  error=None
output='oops'   -> status=completed  error=None
output=5        -> status=completed  error=None

This is already a known flaw — in this file

validate rejects a non-mapping output, and its comment names the execute-side bug outright:

output = config.get("output")
if output is not None and not isinstance(output, dict):
    # execute() silently coerces a non-mapping output to {}, so the
    # author's declared aggregation keys would vanish with no error.
    # Reject at validation, mirroring the command-step (#3262) fix.

The engine does not auto-validate before execute (see WorkflowEngine.load_workflow), so on an unvalidated run that comment describes exactly what happens. This PR closes the execute-side half.

Fix

Fail the step with validate()'s own message, mirroring the wait_for guard a few lines below in the same method:

output_config = config.get("output")
if output_config is None:
    output_config = {}
elif not isinstance(output_config, dict):
    return StepResult(status=StepStatus.FAILED, error=..., output={"results": []})

This is the same "silent empty result + COMPLETED" class the wait_for guards in this method already reject, and the same shape as the fan-out non-list items guard.

No breaking change. An explicit output: (YAML null) stays valid, matching validate — pinned by a second new test. A mapping behaves exactly as before. The only inputs whose behaviour changes are ones that silently produced an empty aggregation.

Verification

• Fail-before / pass-after: the seven parametrized cases fail on unpatched src and pass with the fix.
• tests/test_workflows.py scoped regression: failure set identical to the clean-main baseline captured on 81bf741 (20 pre-existing in this file, all Windows symlink-privilege).
• uvx ruff@0.15.0 check src tests → clean

Tests go into the existing TestFanInStep, beside test_execute_non_list_wait_for_fails_loudly — the sibling guard whose docstring describes the same wiring bug.

---

Written with assistance from Claude Code. Bug found, reproduced, and verified by me on current main.