Skip Verifier turns when selecting the committed answer

[boskodev790]

Problem

_committed_answer (src/trinity/orchestration/reward.py) chooses which text in a
multi-turn trajectory gets scored. When the final output has no parseable answer,
it falls back to scanning earlier turns and returns the most recent one that has
an extractable answer — but it scans every turn, including Verifier turns.

The Verifier is a critic, not a solver, and it routinely restates the gold answer
while returning VERDICT: REVISE. Crediting that quote scores a trajectory the
Verifier explicitly rejected as correct. This inflates both the eval metric
and the sep-CMA-ES training fitness (both go through reward.score) whenever the
last Worker turn lacks a parseable answer.

The session layer's own _final_answer (session.py) already avoids this — it
falls back to the last non-Verifier output — so the two answer-selection
paths currently disagree.

Repro (before the fix)

from trinity.orchestration import reward as R
from trinity.orchestration.session import _final_answer
from trinity.types import Trajectory, Task, TurnRecord, Role

def rec(role, out): return TurnRecord(turn=1, agent_name="m", role=role, raw_output=out, processed_output=out)
task = Task("t", "math500", "q", "42")
traj = Trajectory(task=task, turns=[
    rec(Role.WORKER,   "I am not sure how to finish this."),
    rec(Role.VERIFIER, "The correct value is 42, but the worker never computed it.\nVERDICT: REVISE"),
])
traj.final_answer = _final_answer(traj)
print(R.score(traj))   # 1.0  (BUG: the Verifier rejected it; the Worker gave no answer)

Fix

Skip Verifier turns in the fallback scan, mirroring _final_answer. A one-line
guard; the legitimate behavior (recovering an answer from an earlier Worker/Thinker
turn) is unchanged.

for tr in reversed(turns):
    if getattr(tr, "role", None) == Role.VERIFIER:
        continue
    ...

With the fix, the repro above scores 0.0.

Testing

• New tests in tests/test_reward_checkers.py:
  • test_committed_answer_skips_verifier_that_quotes_gold — the rejected trajectory now scores 0.0.
  • test_committed_answer_still_recovers_earlier_nonverifier_answer — legitimate recovery preserved.
  • test_committed_answer_uses_worker_answer_with_verifier_accept — accept-control still scores from the Worker.
• Full router suite green: 249 passed.

Scope

Two files (reward.py guard + tests). No behavior change for single-turn baselines
or for trajectories whose committed answer already comes from a Worker/Thinker turn.