Add strict proof-completion evaluator contract

[munimthahmid]

What changed

• Reuse the manifest and proof-marker code from proof-from-scratch instead of adding a second copy.
• For marked tasks, use manifest.json to give the agent exactly the supporting files listed for that task and keep them read-only. The agent can edit only the marked proof, and the checker rejects any other changes before grading.
• Add separate prompts for regular and one-shot agents so both follow the new proof boundary.
• Keep the current proof-completion tasks working even though they do not have markers or a manifest yet.

Scope

This PR only adds evaluator and checker support. We can keep the legacy fallback and prompts for now so the current dataset still works. After the generator is merged and we regenerate the dataset with markers and manifest.json, we can remove them.

Testing

• All 1,080 tests pass, including regressions for direct marked-task checks and legacy benchmark dependencies.
• The rebuilt standalone checker passes the real proof-completion and proof-from-scratch boundary tests.
• Three live proof-completion tasks pass with gpt-5.6-sol at medium reasoning.

[munimthahmid]

@Qian-Cheng-nju can you review the PR?

[Qian-Cheng-nju]

LGTM! Thanks a lot!