[P1] Final-state verification independent of tool trajectory

[cursor]

Summary

Introduce a StateVerifier (name flexible) so task success is judged by expected final state, not a single gold tool trajectory. Different valid call sequences that reach the same state should pass.

Priority

priority:p1 — required for Milestone 8 unified evaluation quality (pairs with #76 / #77).

Suggested labels

enhancement, priority:p1, area:probes, area:cli, research

Acceptance criteria

• Probe/task schema supports expected_state (deterministic key/value checks preferred)
• Evaluation tracks tool-call correctness, trajectory correctness, and final-state correctness separately
• Partial success can be represented (some state keys met, others not)
• Failures name which expected state condition did not hold
• Wired into eval report when present; offline/fake tests only in CI
• Docs: docs/final-state.md (or probes.md section)

Non-goals

• Requiring exact gold trajectories
• Auto-executing destructive MCP tools unless explicitly enabled

Related

• Add unified tool-semantics eval command and report #76 unified eval, Add compatibility scorecard and human-readable change explanations #77 scorecard, Implement single and batch agent trace replay #84 trace replay