MCP guard + engine fix: enforce category-scoped operator listing, fix…

[HeRunming]

… run/init param plumbing

Backend

• list_operators (operation_id="list_operators", route /by_category) now REQUIRES a category argument; missing/invalid arg returns a structured ValidationBizError with valid_categories + did_you_mean (difflib) + a three-step recovery hint, instead of returning the full ~145-operator catalog and overflowing the agent context.
• list_operator_categories now returns {count, use_for, not_for, examples} per category (new app/services/operator_category_guide.py), distilled from observed agent failure modes (e.g. text_sft.not_for explicitly routes QA-generation to core_text). Methods without any construction skill inherit operator-selection guidance directly from this response.
• Frontend operator palette keeps unrestricted access via a new list_operators_all (route /, NOT MCP-whitelisted).
• dataflow_engine: drop run/init params the operator's signature can't accept (e.g. agent-supplied system_prompt on operators that don't take it) instead of crashing with TypeError; unwrap llm_serving={"id": ...} dict to its scalar id (was crashing with unhashable-type).

Skills

• Promote .claude/skills/{dataflow-dev,dataflow-operator-builder, generating-dataflow-pipeline,prompt-template-builder} from the stale "* copy/" siblings (renames, not duplicates).
• .gitignore now excludes .claude/skills/* copy/ to prevent recurrence.

Evaluation impact (12 tasks x 3 reps x 4 agent configs, see emnlp paper):

• Mean operator errors: 1.87 -> 0.24 across the four agents (7.8x).
• Context-overflow failure rate: 5.7-21.7% -> 0% on all four conditions.
• End-to-end execution success: best agent rises from ~16% to 47.2%.