feat(docs): kbagent docs query — answer questions from Keboola documentation (port docs_query)

[padak]

Part of #390.

Summary

Add a native kbagent docs query "<question>" command that answers questions from the Keboola documentation, porting the docs_query MCP tool.

MCP source (keboola-mcp-server @ agent-v1.60.0)

• Tool: src/keboola_mcp_server/tools/doc.py (docs_query) — thin, read-only; delegates to the AI-service client.
• Client: clients/ai_service.py docs_question(query) → POST docs/question, body {"query": "<str>"}, returns {"text": str, "sourceUrls": [str]} (wire field is camelCase sourceUrls).
• Base URL: https://ai.{stack-suffix}; auth header X-StorageAPI-Token.
• It is a server-side RAG/Q&A endpoint — no local retrieval, no embedded corpus.

What kbagent already has

• ai_client.py AiServiceClient already targets the same ai.{stack} base URL (_derive_service_url(stack_url, "ai")) with header X-StorageApi-Token, and inherits retry/backoff/timeout from BaseHttpClient.
• It already implements get_component_detail (GET /docs/components/{id}) and suggest_components (POST /suggest/component) — so the camelCase→snake_case alias pattern is established.
• Missing: one client method docs_question(), a service method, a thin command. No new client, no new auth, no URL work.

Proposed command

kbagent docs query "How do I configure incremental loading?" [--project ALIAS]

• New docs Typer group (room for a future docs suggest-component, since the /suggest/component client method already exists with no command).
• --project optional (resolve stack+token like component detail).

Implementation sketch (3-layer)

• L3 ai_client.py: docs_question(query) -> dict → POST /docs/question, body {"query": query}.
• L2 add ask_docs(alias, query) to ComponentService (it already owns ai_client_factory) or a small DocsService.
• L1 commands/docs.py: thin command, dual output (JSON {text, source_urls}; human = answer panel + "Sources" list). Add DocsAnswer model to models.py with sourceUrls alias.

Effort: S

Zero new infrastructure; one client method + one service method + thin command + model.

Acceptance criteria

• kbagent docs query "…" returns answer text + source URLs.
• --json emits {text, source_urls, project_alias}.
• Unit tests (mock AI client) + E2E test.
• Doc-sync surfaces per Epic: Port keboola-mcp-server agent tools into native kbagent commands #390 checklist.

Risks / open questions

• docs/question is an internal AI-service endpoint with no published contract found in either repo — confirm shape stability with the AI Service team (response model breaks silently if text/sourceUrls rename).
• Requires a project Storage token (no global/anonymous mode), even though docs are global content.
• Redundant with kbagent tool call docs_query, but native is faster (no subprocess) and gets --json/Rich.
• Consider shipping docs suggest-component in the same PR for symmetry (client method already present).