[Cookie0103]

A dependency-ordered study can currently run children before parents in the sequential backend, submit part of an invalid Ray graph, and associate Ray results with the wrong experiment when references finish out of order. Shared agent resources can also be prepared repeatedly or left open after partial preparation and cleanup failures.

This repairs the experiment execution lifecycle across graph validation, dispatch, monitoring, and resource cleanup:

• Validate IDs, missing dependencies, cycles, and episode budgets before preparing agents or experiment directories; use a stable, iterative topological order for sequential and Ray execution.
• Keep result identity attached to each Ray reference and enforce each episode's own runtime budget. Dependency/worker queue time is excluded, and cancellation escalates once after a monotonic grace period.
• Prepare and close shared agent objects once, attempt all cleanup callbacks, preserve the primary failure, and clean partially initialized Ray runtimes only when owned by this invocation.
• Reject dependency graphs on joblib before side effects; retain its independent-experiment behavior.

Validation on the frozen lock (Python 3.12.13, Ray 2.48.0, pytest 7.3.2): 133 passed, 1 skipped, 6 deselected in the experiment suite. The 120 new cases cover deterministic failure regressions, native ExpArgs persistence/relaunch, and real local Ray execution, including failed-parent isolation and per-episode deadlines. The independent review rerun passed all 112 deterministic cases. Black 25.1.0, darglint, compileall, and diff checks pass.

Browser/provider-dependent tests were excluded; no real LLM or remote benchmark was exercised. Ray runtime deadlines still depend on task-state availability; sequential/joblib deadlines remain unenforced. These limitations and the changed joblib/shared-agent behavior are documented.

Prepared with AI assistance.