feat(lark-cli): Feishu approval submission via official lark-cli

[an7tang]

Overview

Adds a built-in lark_cli capability that submits Feishu (Lark) approval requests on behalf of the user through the official larksuite/cli, together with a lark_approval_submit SOP card (search definition → assemble form → dry-run preview → user confirmation → real submission).

Key design

• Layered command policy (app/lark_cli/policy.py): permissive reads, strict writes — registered read-only commands accept the CLI's native flags as-is; unregistered commands are adjudicated dynamically via the CLI's own Risk: classification; write commands are limited to a strict whitelist; the raw api escape hatch, --yes, and caller-supplied uuid are permanently blocked.
• Fully in-conversation setup: app credentials can be configured in chat (config init, or --new to create an app on the spot); user login uses the multi-turn device-code flow.
• Submission confirmation gate: a real submission must match a dry-run preview already persisted in the current session (submission_digest); the trusted layer replays the confirmed content and injects --yes plus an idempotent uuid. Form structure is validated against the approval definition before preview.
• Errors as guidance: frequently misleading errors (auth scopes lacking admin permissions, auth qrcode unavailable, server-side form rejections, etc.) get deterministic corrective guidance appended by the trusted layer, preventing the model from misreading them as task failure.
• Minimal engine-side support: built-in capabilities can now be marked as required by SOP nodes; when resuming a paused frame, the latest user reply and the current time (user timezone) are injected.

Testing

• New backend/tests/test_lark_cli.py (policy + end-to-end service tests against a fake CLI binary).
• After rebasing onto the latest main, the full backend suite passes (2078 tests); ruff reports zero new issues in the changed files.