Release MCP

Review every release. Inherit scope and session authorization from Release a New Version.

1. Compare product changes with apps/cli/src/mcp.rs, crates/api-cloud/src/mcp.rs, crates/mcp, shared crates/agent-access, and current contract tests. Review initialization, tools/resources, schemas, pagination, errors, capability exposure, and local/hosted differences. Derive expected tools from current source, never a remembered tool count.
2. Preserve local proposal-approval and hosted read-only boundaries, authentication, scopes, and workspace isolation. Implement required coverage before candidate freeze or document the intentional difference/explicit deferral. Do not infer supported agent behavior from the database schema alone.
3. Run affected local and hosted contract/auth tests from cli_ci.yaml and api_ci.yaml, including allowed and denied paths. Reuse the CLI test evidence rather than rerunning identical shared checks.
4. For each shipped desktop platform, use the packaged CLI and isolated fixture database for a real stdio initialize/discovery exchange and affected calls. Keep stdout protocol-only and verify clean shutdown. Reuse the artifact and verify-packaged-cli.py evidence from Release CLI.
5. Publish required hosted changes through Release API. Verify the serving revision, authentication discovery/OAuth resource metadata, and authenticated discovery plus an affected operation with an authorized test account. Health/401 responses or mocks are not hosted runtime acceptance. Keep writes confined to supported local disposable fixtures and approval rules.
6. Review shared host manifests/setup references once via Agent packages. Update public setup instructions through Release Docs.