Probe Should Discover Models, Not Spend Money

The Broader Goal at the End of the Previous PR

Make diagnostic commands safe to run routinely while the launch-credibility review closes correctness and latency gaps.

Strategic Shifts in This PR (if any)

Probe is explicitly diagnostic. Discovering a harness/model pair no longer implicitly starts paid calibration work or dispatches agents while a database write transaction is open.

What This PR Shipped

The probe records a newly discovered pair in harness_models with active status and self_report discovery source, then commits and returns. It emits one concise instruction to run synlynk capability sweep when calibration is wanted. Regression tests cover the import and dispatch boundary, persisted metadata, and release of the SQLite write transaction.

Brainstorm Visuals Used

None.

What This Achieved on the Path to Autonomy

Routine health checks are now bounded, non-spending observations. Calibration remains an intentional operator action, while uncalibrated models can still participate through the routing explore bonus.

Strategic Note: The Goal at the End of This PR

Diagnostics should remain fast and side-effect constrained; expensive fleet actions belong behind explicit commands and their own operational budget.