Extract standalone code intelligence for agents and teams

[eren23]

Code intelligence currently ships inside the coding agent, which makes it difficult for Codex, Claude Code, Cursor, and team services to use independently. This extracts it into attocode-code-intel 0.1.0 with native MCP onboarding, while preserving legacy imports and freezing the agent surface.

• A shared operation gateway serves local stdio and authenticated Streamable HTTP/REST, with workspace isolation, freshness checks, bounded responses, and daily/full tool profiles.
• Local knowledge stays in SQLite; team knowledge uses repository-scoped PostgreSQL with authorship, revision anchors, and explicit sharing. Remote analysis uses committed Git worktrees and checks authorization before cache access.
• Query analysis snapshots the AST cache so background hydration cannot invalidate iteration; a deterministic concurrent-write regression covers this.
• The dashboard adds agent connection instructions and knowledge editing. Standalone Docker, package, migration, performance, and publishing workflows support the new distribution.

Validation: 2,066 existing regression tests passed (5 skipped, 1 expected failure); 24 standalone product tests passed; fresh PostgreSQL migrations, standalone wheel isolation, frontend build, and Docker runtime smoke passed locally. Actual Codex CLI 0.153.4 and Claude Code 2.1.212 sessions called the installed wheel's symbols tool successfully. Cursor configuration and protocol checks passed; an actual Cursor agent session remains unverified.

The synthetic 10,000-source-file evaluation recovered all 9,999 dependency edges, with 7.32 s cold bootstrap and 246 ms focused lookup on macOS. Workload limits and detailed evidence are in docs/intelligence-validation.md and packages/code-intel/evals/results/.

Publishing uses intel-v* tags and the pypi-intelligence trusted publisher, gated by the complete intelligence workflow. PyPI account setup remains a release prerequisite; this change does not deploy a hosted service.