[FIX]: Evaluate probe verdicts over final traces

[Spencer Schoenberg (spencrr)]

Description

Changes behavioral probes to evaluate the verdict once over the completed trace. Probes no longer stop on a detected verdict by default; callers can provide stop_when when an online condition should define an earlier terminal trace.

This fixes sequence and prior-condition probes that cannot be resolved correctly from partial prefixes. Terminal verdict evidence is stored on Result.terminal_evaluation, online stop evidence remains on turns with EvaluationPurpose.STOP_CHECK, and trace completion is recorded with TraceEndReason. Terminal evaluation stays inside the active session.

This is the first layer that changes verdict cadence, so it bumps the private xdist envelope from v2 to v3. Controllers reject v2 envelopes rather than combining prefix-derived and terminal-trace statuses.

Depends on #148. Because the branches live on a fork, this PR temporarily includes lower-layer diffs and targets main; those diffs disappear as dependencies merge.

Breaking changes

Behavioral change: Result.eval_results remains a turn-only compatibility view and is normally empty for probes without stop_when; use Result.terminal_evaluation for terminal verdict evidence. Adaptive drivers without an online stop receive no evaluator feedback and may run to max_turns.

Checklist

• pre-commit run --all-files passes
• Tests cover one-call verdict cadence, sequence and prior-condition probes, scopes, stopping, feedback, budgets, session ordering, summaries, and error results
• Documentation updated

Validation after reslicing: the full suite passes with 1,178 tests and 5 skips; all pre-commit checks pass.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.