fix(cli): preserve unresolved sandbox failures across unrelated tools

[jsiu93]

Summary

Track the failed tool name with each unresolved sandbox failure. A later success now proves recovery only when it comes from the same tool in a different Step. An unrelated tool_search leaves a failed Read unresolved, so non-interactive boundary denial returns exit code 1 with empty stdout.

Fixes #4388

Verification

mise exec node@24.19.0 -- npx -y npm@11.19.0 --workspace maka-agent run build
mise exec node@24.19.0 -- node --test --test-name-pattern "keeps a sandbox failure unresolved" packages/cli/dist/__tests__/runtime-host-run-command.test.js
Before fix: 0 pass, 1 fail, exit 1

Before the fix, the failing assertion received exit code 0; it expected 1.

mise exec node@24.19.0 -- node --test --test-name-pattern "sandbox failure|sandbox boundary|boundary request" packages/cli/dist/__tests__/runtime-host-run-command.test.js
After fix: 11 pass, 0 fail, exit 0

mise exec node@24.19.0 -- node --test packages/cli/dist/__tests__/runtime-host-run-command.test.js
36 pass, 0 fail, exit 0

A live CLI run against the patched build returned 1 after 21 seconds, kept stdout empty, and retained the non-interactive sandbox boundary diagnostic on stderr.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: pi (gpt-5.6-sol), human-reviewed

Checklist

• Tests cover the change, with a red test recorded before the fix
• The CLI build and affected test file pass locally

Does this PR entail a change in behavior?

• Yes, described under Summary above
• No