feat(cursor): add first-class model, readonly, is_background subagent frontmatter fields

[dyoshikawa]

Summary

Follow up Cursor upstream updates by promoting the documented Cursor subagent frontmatter fields to first-class, typed optional fields on CursorSubagentFrontmatterSchema:

• model (string, default inherit)
• readonly (boolean, default false)
• is_background (boolean, default false)

Previously these three fields already round-tripped through the cursor: section via z.looseObject passthrough, so there was no data-loss bug — this change makes them first-class so they are explicitly typed and validated (e.g. wrong types like readonly: "yes" are now rejected).

The schema keeps z.looseObject, and toRulesyncSubagent/fromRulesyncSubagent continue to spread the fields into/out of the namespaced cursor: section, so no processor wiring change was needed. Cursor subagents remain project-scope only (.cursor/agents/); no global scope was added.

Out of scope (intentional)

The issue also mentions a canonical top-level model mapping. This is intentionally skipped: rulesync has no canonical top-level subagent model, and each tool keeps model in its own namespaced section (e.g. claudecode.model, opencode.model, kilo.model). Adding a canonical mapping would contradict that existing convention.

Evidence

• https://cursor.com/docs/agent/subagents
• https://cursor.com/changelog/2-4

Changes

• src/features/subagents/cursor-subagent.ts: extend schema with model, readonly, is_background.
• src/features/subagents/cursor-subagent.test.ts: add happy-path tests for parsing the fields as first-class and round-tripping them through the cursor: section (file import + generate), plus an invalid-type rejection test.
• docs/reference/file-formats.md and skills/rulesync/file-formats.md (auto-synced): document the cursor: subagent block.

Verification

pnpm cicheck is fully green (code: fmt/oxlint/typecheck/267 test files, 6422 tests passing; content: skill-docs sync, cspell, secretlint all passing).

Closes #1882

🤖 Generated with Claude Code

[dyoshikawa]

@dyoshikawa Thank you!