fix: route sync output_invalid through _fail_run

[rickgorman]

Summary

Sync executor.start() results that fail declared-output validation (artifact is empty / missing fields) currently mark the run FAILED and call _halt_job immediately. The async tick path already routes the same failure through _fail_run(..., error_category=\"output_invalid\"), which is what lets exit rules retry.

This made a loop on output_invalid fire for async executors and never fire for sync ones (script / callable). Arena hits this on stepwise-run --local --wait with a script invoke that must honor output contracts.

Reproduce

git clone https://github.com/zackham/stepwise.git
cd stepwise
uv sync
uv run pytest tests/test_engine.py::TestSyncOutputInvalidExitRules -q

On current master (this PR's parent):

FAILED tests/test_engine.py::TestSyncOutputInvalidExitRules::test_missing_output_retries_via_exit_rule
AssertionError: expected retry-then-complete, got failed; calls=1 runs=['failed']

The test is a callable step that returns {} on attempt 1 (declared output status missing) and {\"status\": \"ok\"} on retry, with an exit rule:

ExitRule("retry_invalid", "field_match", {
    "field": "error_category",
    "value": "output_invalid",
    "action": "loop",
    "target": "emit",
    "max_iterations": 5,
}, priority=10)

With this change the same test completes after two runs: FAILED (output_invalid) then COMPLETED.

Minimal CLI equivalent (same engine path as --local --wait):

# /tmp/invalid-then-ok.yaml
name: invalid-then-ok
steps:
  emit:
    executor: script
    run: |
      python3 -c 'import os,json,pathlib; p=pathlib.Path("/tmp/sw-attempt"); n=int(p.read_text()) if p.exists() else 0; p.write_text(str(n+1)); print("{}" if n==0 else json.dumps({"status":"ok"}))'
    outputs: [status]
    exit:
      - name: retry_invalid
        type: field_match
        priority: 10
        config:
          field: error_category
          value: output_invalid
          action: loop
          target: emit
          max_iterations: 5

rm -f /tmp/sw-attempt
uv run stepwise run --local --wait /tmp/invalid-then-ok.yaml; echo exit=$?

Unpatched: job fails after one empty artifact. Patched: retries and completes.

Change

In _process_launch_result for result.type == "data", when _validate_artifact / _check_artifact_size fails, call _fail_run with error_category="output_invalid" instead of _halt_job. Same as the async poll path a few hundred lines above.

No exit-rule change; jobs without a matching rule still halt via _fail_run's existing fallthrough.