fix(swarm): block downstream tasks when upstream fails

[omcdecor-cyber]

Summary

When a task fails in the DAG, downstream tasks that declare it as a
dependency are dispatched anyway, with an empty upstream context. This is
most visible in the investment_committee preset, where a failed
risk_officer lets portfolio_manager produce a "decision" with no risk
input — clearly unsafe for any production use.

The layer-boundary _sync_run_tasks_snapshot added in #132 surfaces task
failure to polling clients sooner, but it does not gate dispatch — PM
still runs the moment the layer containing the failed risk task finishes.

Root cause

• _execute_run flips all_succeeded = False on task failure but never
  breaks out of the layer loop or gates the next layer.
• The downstream worker's upstream-summary loop in _execute_layer reads
  from task_summaries via if source_task_id in task_summaries — a
  failed upstream never populates that dict, so the lookup silently
  yields no upstream context. The worker then runs with whatever its
  prompt template makes of an empty {upstream_context}.

Fix

_execute_layer: before submitting each task, walk task.depends_on
and load each upstream from TaskStore. If any upstream has
status != completed (failed / blocked / cancelled / missing), mark this
task TaskStatus.blocked, record blocked_by + reason, emit
task_blocked, and skip executor.submit. Same-layer peers with no
shared upstream are unaffected — the loop continues.

_execute_run: blocked tasks are absent from layer_results. After the
result-processing loop, any layer_task_id missing from layer_results
was blocked → set all_succeeded = False so the run is finalized as
RunStatus.failed.

Blocked state cascades naturally through deeper layers: layer-3 task Z
depending on a blocked layer-2 task Y will see Y.status="blocked" != completed and block in turn.

Test plan

• New file agent/tests/test_swarm_dag_gating.py with 3 regression
  tests, all TDD red-green verified against this diff:
  • test_failed_upstream_blocks_downstream — PM is blocked, not
    failed or completed, when risk fails
  • test_blocked_downstream_emits_task_blocked_event — events.jsonl
    contains task_blocked for PM, NOT task_started
  • test_run_marked_failed_when_downstream_blocked — run finalizes as
    RunStatus.failed
• Full swarm test suite passes (115 tests in -k swarm, 0 regressions)
• Manual replay against the canonical 2-task DAG (risk → pm) confirms
  events.jsonl post-fix shows task_blocked not task_started for PM,
  and PM's started_at remains None.

Diff is +181 LOC additive (no behavior change for previously-passing DAGs).