v0.3.5

Fixed

• Keep cold-start benchmark priors separate from measured outcome history.
• Refresh Pi/OMP catalog and Codex capacity during long-running sessions.
• Learn model quality only from explicit failures and verified outcomes.

Added

• Add opt-in Pi/OMP routing through switchyard/auto, including OMP prewalk handoffs.
• Add commands to mark the last routed run as verified successful or failed.