Compare ABC-VLA and adapted pi0.5 in repeatable bimanual episodes

[niksacdev]

ROVE previously compared individual action chunks without running the robot task to completion. This adds an experimental ABC bimanual executor that compares ABC-VLA and an ABC-adapted pi0.5 checkpoint through complete observe–predict–execute episodes, using the existing Trials, Campaigns, SQLite evidence, traces and reliability reports.

Cases fix simulator resets; policy seeds supply repeated attempts. Both strategies use the same task, camera/action contract, execution budget and explicit success criterion. Reports distinguish simulator outcomes, worker failures, policy-call timing and unmatched initial states. A focused UI and equivalent CLI require review and registered local runtimes before launch.

A separate prepared ABC/XDOF-to-LeRobot converter and explicit pi0.5 training recipe preserve training splits, checkpoint lineage and normalization. Strict checkpoint loading prevents an upstream silent fallback to uninitialized weights. Dependencies remain isolated; timeout/cancellation stops the entire episode process group.

Validation includes actual LeRobot conversion/readback of two training episodes (1,870 frames), and real ABC CPU simulation with rendered cameras, 30 control steps and durable trajectory evidence. The simulator smoke used a synthetic hold-position policy. GPU training and learned-policy comparison have not run, so this PR claims integration support, not model performance. Product specification, setup guide, diagrams and ADR-038 describe that boundary.

All configured local pre-commit hooks passed, including the full Python suite, Ruff, formatting, Bandit, secret scan and Markdown validation. All 227 frontend regression tests passed; setup/review screens were visually inspected. Nested process cleanup and corrupted evidence/checkpoint failures have direct regression coverage.

Closes #34.