Harden graph runtime, durable runs, and plugin portability

[JJPuertas]

Summary

• fix Mermaid output, global END observability, immediate-END lifecycle events, strict deadline accounting, and per-node state isolation
• harden detached runs with PID liveness, cross-process resume claims, stale-result protection, restart-safe event sequencing, run-id validation, and bounded Windows sharing retries
• enforce unambiguous condition rules and strict JSON boolean approval while safely serializing workspace-editing backends
• make the MCP launcher work under both Claude Code and Codex on Windows, move large CLI prompts off argv, propagate model/timeouts, and extract the Codex final message cleanly
• update plugin documentation and add focused regression coverage for every repaired behavior

Why

The initial v0.1 review found several small but user-visible correctness and portability failures: Mermaid diagrams could not render, the plugin launcher depended on python3 on Windows, two MCP hosts could misclassify or double-resume shared runs, coding backends could execute concurrently when cwd was omitted, and malformed reviewer/condition output could approve or strand a graph incorrectly. This change closes those issues without adding runtime dependencies.

Impact

Diagrams render correctly, detached runs behave consistently across simultaneous Claude/Codex hosts, coding-agent defaults protect shared workspaces, reviewer routing is fail-safe, long prompts avoid the Windows command-line limit, and the bundled plugin cold-starts through one shared configuration on both hosts.

Validation

• uv run --directory plugins/athena-graphs --extra dev --extra mcp pytest ../../tests -q: 95 passed
• python -m compileall -q agentgraph: passed
• git diff --check: passed
• Codex plugin validator: passed
• claude plugin validate plugins/athena-graphs --strict: passed
• Claude Code and Codex MCP cold starts: passed with tool registration
• live Codex stdin plus --output-last-message transport probe: passed

Tested on Windows. macOS/Linux CI remains a follow-up portability check.