fix(runner): persist resumed tool guardrail results

[FU-max-boop]

Summary

This pull request fixes a recovery gap after resolving interrupted function-tool approvals with a client-managed Session. When an approved tool, its input and output guardrails, and its lifecycle hooks completed but the resumed Session append raised, the caller retained a RunState that omitted the completed guardrail audit results. A later retry could reconcile the tool call and output through #4650, but it could not reconstruct those results without rerunning user code.

• Publish the accepted guardrail-result prefix plus the current resumed step's results to the same RunState before the fallible Session append in both runner paths.
• Keep caller-mutated streaming result lists outside the durable ownership boundary by sourcing the streaming prefix from accepted state.
• Preserve the existing public result types, ordering, and RunState schema.

Test plan

• Added a four-row public Runner matrix covering same-mode and cross-mode recovery, live and JSON state, and both before-commit and commit-then-raise Session outcomes.
• Verified exactly-once tool effects, input/output guardrails, start/end hooks, Session history, public results, durable state, and final model input.
• Ran .agents/skills/code-change-verification/scripts/run.sh on the final commit and current main; formatting, lint, typecheck, and the full test suite all passed.

Issue number

Closes #4646

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR