Add a reproducible multi-agent orchestration pilot

[vitormattos]

Summary

Implement the remaining executable dependency for LibreSign/libresign#8989 without introducing another skill or orchestration framework.

The OpenAI Agents API now provides real independent subagent contexts. This PR extends the existing execute-issue evaluation path to use that capability for a controlled comparison.

It adds:

• an execute-issue reference describing the multi-agent pilot protocol;
• a text-only/read-only runner that executes the exact same task packet twice:
  • single-flow, with no multi_agent configuration;
  • coordinated, with multi_agent.enabled=true and a small explicit subagent limit;
• recording of wall-clock time, subagent creation events/IDs, output, raw event types and usage when the runtime exposes it;
• explicit manual fields for correctness, interventions, duplicated work and handoff failures rather than inventing those metrics;
• a shareable synthetic fixture containing two genuinely independent investigation questions;
• executable unit coverage proving request parity between the two modes.

Scope

This is deliberately not:

• a new skill;
• a permanent role-agent system;
• a new workflow/spec format;
• tied to Behat;
• tied to Docker/NCDD;
• enabled in normal CI.

The initial comparison uses environment.type=none so that the first orchestration experiment measures coordination rather than repository/runtime differences. Repository execution can be evaluated separately later.

Validation

• three incremental commits created with DCO and verified cryptographic signatures;
• runner syntax and request-construction test executed locally;
• Validate agent skills / package is green, including unit/package tests and REUSE licensing.

Relates to LibreSign/libresign#8989.