Agent workflow receipt claims `external_api_touched=true` when the agent made zero tool calls on station-v0.78

Reproduction steps:

1. Use a clean RailCall Station v0.78 workspace.
2. Run one type="agent" node with a tool allowlist and a zero-spend budget:

{
  "id": "a",
  "type": "agent",
  "tools": ["fake_read"],
  "budget": {"max_steps": 2, "max_tool_calls": 1, "max_spend_cents": 0}
}

1. Inject an LLM response that returns a final answer on the first turn with tool_calls=[].
2. Run the real workflow_engine.run_workflow(..., allow_live_effects=true).
3. Pass the resulting node receipt to the real receipt_summary.summarize() helper.

Expected:

• tool_calls=0, action_leaves=[], and spent_cents=0 mean no integration was planned or applied.
• The receipt must say external_api_touched=false.
• The summary must say that nothing left the station.

Actual:

workflow outcome     = "COMPLETED"
tool_calls            = 0
spent_cents           = 0
action_leaves         = []
external_api_touched = true

The summary then reported:

Touched: 1 actions · an internal step

Control result:

• The same no-tool agent with allow_live_effects=false reported external_api_touched=false.
• No provider call, socket, credential, or external write occurred in either run.

Root cause:

• workbench/workflow_engine.py:962-969 sets action receipts' external_api_touched from bool(allow_live_effects).
• workbench/workflow_engine.py:997-1003 sets the node-level agent receipt from the same flag, without checking tool_calls, action_leaves, or the actual integration mode.
• workbench/primitives/receipt_summary.py:94-123 treats external_api_touched as the fact used to render the “Touched” line.

Impact:

A signed, integrity-valid receipt and its standard summary can falsely attest that an external effect occurred when the agent produced only a local answer. This corrupts operational/audit interpretation of whether data or an effect left the station. No external request occurred in the reproduction.

Suggested fix:

Derive external_api_touched from the action receipts that actually executed in live mode. An agent with no tool calls or only mock/pure actions must produce false.