feat: add assertions schema to EvalCriteria for code-based grading

[melmennaoui]

Summary

Add code-based assertion support to the evaluation criteria schema, enabling declarative grading checks that run against agent output without requiring an LLM judge.

Changes

pkg/session/session.go

• Assertion struct — Name, Type, Value fields for declarative checks. Supported types: contains, not_contains, equals, starts_with, ends_with, regex, cost_threshold, tool_called.
• Assertions field on EvalCriteria — optional list of code-based assertions.
• Verify field on EvalCriteria — optional shell script for post-agent outcome verification.
• AssertionsCheck — result type with passed/total counts and per-assertion results.
• AssertionResult — individual assertion outcome (name, type, passed, reason).
• VerifyCheck — verify script result (passed, exit_code, output).
• Added both to EvalResultChecks.

pkg/session/branch.go

• Updated cloneEvalCriteria to deep-clone the Assertions slice.
• Updated cloneEvalResultChecks to deep-clone AssertionsCheck and VerifyCheck.

pkg/session/session_test.go

• Added test cases for JSON unmarshaling of assertions and verify fields.

Example eval JSON

{
  "evals": {
    "relevance": ["The agent created the file"],
    "assertions": [
      {"name": "mentions file", "type": "contains", "value": "hello.txt"},
      {"name": "used write tool", "type": "tool_called", "value": "write_file"},
      {"name": "cost under budget", "type": "cost_threshold", "value": "0.50"}
    ],
    "verify": "test -f /workspace/hello.txt"
  }
}

Closes #4082