feat(rubrics_verifier): grade a multi-turn run as of turn k

[kaihackney-scale]

To chart how a simulated-user conversation's score climbs, each turn's env snapshot (prompt_agent's snapshot_config.per_turn) has to be graded on what the agent had done by then. The rubric verifier always loads every turn, so a criterion about what the agent said would be credited from a later turn. grade_up_to_turn (1-based, also settable per run through step_params) makes the judge see only turns 1..k: the per-turn trajectories, the turn-k trajectory as the single-trajectory fallback, and the user's per-turn messages are cut off there. The stored response and structured_output are the conversation's last reply, so the judge instead gets a note that the reply at this point ends turn k's trajectory. The verification entry records graded_up_to_turn. The run's stored context is never modified, and with it unset, or k at or past the last turn, grading is unchanged.

Validation: new unit tests run the verifier on a three-turn run (turns 1..2, turn 1 alone, the per-run override, k at the last turn, unset, invalid values, round-trip); all but the unset case fail without the change. All verifier unit tests and ruff pass. The full unit suite's other failures (test_mcp_cli_builder, a parallel-only flake in task_run_teardown_test) also occur on main locally.

🤖 Generated with Claude Code

Confidence Score: 5/5

The PR appears safe to merge; both previous findings are fixed.

Summary

Adds grade_up_to_turn so rubric grading can use only the conversation through a chosen turn.

• Rubric verification can score a conversation as of turn k.

Both previous findings are addressed. No new actionable issues were found.

_{Reviews (2) · Last reviewed commit: "fix(rubrics_verifier): grade_up_to_turn ..." · Reviewed by Greptile}