Agentic Subprocess: Complete Tool Invocation Only After Entire Activity Chain Finishes

**User Story**
As a process modeler, I want the agentic subprocess orchestrator to treat a startable activity and all downstream activities connected by sequence flows as a single tool invocation, so that the LLM only resumes when the complete fragment has finished executing.

**Background**
Today the orchestrator signals tool completion when the selected startable activity ends. This can result in the LLM resuming while downstream activities in the same process fragment are still executing.

The workflow engine already tracks execution completion correctly across sequence flows, gateways, conditions, and loops. The orchestrator should align its completion signal with the engine's notion of fragment completion.

**Acceptance Criteria**

- Tool completion is not triggered when the startable activity ends if downstream activities remain active.
- Tool completion is triggered only when the entire executable fragment has completed.
- Sequence flows following the startable activity are included automatically.
- Gateways and conditional paths are handled correctly using the engine's existing execution semantics.
- Loops within the fragment are handled correctly.
- The LLM receives exactly one completion event per tool invocation.
- Existing single-activity tools continue to function unchanged.

**Technical Approach**

- Hook into the existing plugin extension point where the engine determines that a child execution has fully completed.
- Use the engine's execution completion semantics as the source of truth.
- Add parse-time validation to ensure agent subprocesses disable engine auto-complete behaviour.
- Fail deployment when incompatible completion settings are detected.

Based on https://github.com/finos/fluxnova-plugins/issues/53

Start at the plugin extension point where the engine determines that a child execution has fully completed, and trace how agentic subprocess tool completion is currently signaled. Then inspect parse-time validation for engine auto-complete settings; done means completion waits for the full executable fragment, handles flows, gateways, conditions, and loops, emits once, and rejects incompatible deployments.

Written by the indexing model from the issue text.