fix(mcp): declare readOnly/destructive/openWorld hints on every listed tool

[ctkm-aelf]

Summary

• OpenAI's plugin portal scan rejects NyxID's MCP server: every listed tool must declare explicit boolean readOnlyHint, destructiveHint and openWorldHint annotations, and tools/list only sent name, description and inputSchema (42 findings across the 14 nyx__* meta-tools).
• tools/list now attaches MCP annotations from a new mcp_service::tool_annotations:
  • Meta-tools use a fixed table: discovery/read tools (search_tools, discover_services, list_connected_services, wait_for_connection, ssh_list_services, oracle_pools/result/session) are read-only; connect_service and oracle_ask/attach/extract are open-world writes; call_tool and ssh_exec are destructive.
  • Per-service tools derive hints from their HTTP method (GET/HEAD/OPTIONS read-only, POST non-destructive write, other methods and generic proxies destructive).
• Response shape is additive only; no stored data, auth, or request handling changes.

Test Plan

• New tests: tool_annotations_cover_every_meta_tool_with_boolean_hints, tool_annotations_follow_service_endpoint_method
• cargo test -p nyxid -- tool_annotations mcp_transport::tests → 76 passed (with local MongoDB)
• cargo clippy -p nyxid --all-targets -- -D warnings clean; cargo fmt clean
• After deploy: rerun the OpenAI plugin portal MCP scan and confirm the annotation findings are gone

Checklist

• Code follows the project's architecture rules
• No hardcoded secrets or credentials
• Error messages do not leak internal details
• Documentation updated (if applicable) — n/a

Follow-up (not in this PR): OpenAI's plugin guidelines disallow "a generic executor to enable operations not individually exposed for review", which nyx__search_tools + nyx__call_tool may trip at human review.

🤖 Generated with Claude Code

[github-actions]

📊 Code coverage

Component	Lines	Threshold	Status	Δ vs base
Backend (nyxid)	87.07%	73%	✅	🔺 +0.02

Gate: line coverage must stay at or above the threshold. Ratchet plan (W21): Backend → 55%, CLI → 50%, Frontend → 30% by quarter end.