Improve SCHEMA_MISMATCH repair guidance and add loop-alias/exports repair tools

[zhongkaifu]

Motivation

• Improve LLM guidance and deterministic tooling to more accurately and efficiently fix SCHEMA_MISMATCH validation errors when repairing workflows.
• Reduce unnecessary LLM rounds by applying deterministic repairs for common loop-related reference issues and missing exports.
• Make repair tools callable from the LLM agent so automated fixes can be applied via the existing toolchain.

Description

• Enhanced the LLM system prompt and SCHEMA_MISMATCH guidance in velvetflow/planner/repair.py to prioritize schema-aware fixes and Jinja-based conversions (e.g., join/first, tojson).
• Added two deterministic repair tool wrappers to the agent in repair.py: align_loop_body_alias_references and fix_missing_loop_exports_items, and included them in the agent tools list.
• Extended velvetflow/planner/repair_tools.py apply_repair_tool to handle align_loop_body_alias_references and fix_missing_loop_exports_items by routing to the existing implementations.
• Kept changes minimal and scoped to repair prompt/text and tool exposure; primary modified files are velvetflow/planner/repair.py and velvetflow/planner/repair_tools.py.

Testing

• No automated tests were executed as part of this change.
• Existing unit tests referencing SCHEMA_MISMATCH remain unchanged and should be run in CI to validate behavior.
• Manual linting/format checks were not performed in this rollout.
• Recommend running the test suite (pytest tests/) in CI to verify regressions and behavior on loop-related repair cases.

---

Codex Task