Proofs: regrade stored runs when only tests/ changed #139

Is your feature request related to a problem? Please describe.

When a check changes after the first proof, the proof runs again from scratch: NOP, Oracle and every negative control, plus any stored agent runs. In an agent-optimization loop on a FreeCAD CAD agent, a run with 8 check changes spent about 226 minutes on proofs (4 to 85 minutes each). Most changes touched only tests/, so the agents' and controls' outputs were still valid; only the grades needed redoing. Harbor already supports this (harbor jobs regrade re-runs the verifier on stored trials). The task-create proof, however, fingerprints tests/ and rejects reused runs, so the skill cannot simply be told to regrade.

Describe the solution you'd like

When only tests/ changed, let a proof regrade stored control and agent jobs instead of re-running them, and record that it did (old and new tests/ hashes, the jobs regraded). Re-run only what the change can affect, for example a new negative control.

Describe alternatives you've considered

Regrading by hand outside the skill, which loses the proof record.

Additional context

Example: NVIDIA/nemoclaw-community#213