fix: prompt Message wrapping + add detail_level=minimal to get_architecture_overview

[selika]

Summary

Two independent fixes, both surfaced from the same MCP slash-command failure (/code-review-graph:architecture_map):

1. fastmcp ≥3.2 prompt Message wrapping (regression fix)

fastmcp 3.2.4 stopped accepting raw dict messages in prompt return values:

McpError: messages[0] must be Message or str, got dict.
Use Message({'role': 'user', 'content': '...'}) to wrap the value.

This breaks all 5 MCP prompt templates (review_changes, architecture_map, debug_issue, onboard_developer, pre_merge_check) since they return list[dict].

Fix: wrap each message via fastmcp.prompts.prompt.Message so role is explicit and future multi-turn prompts compose naturally. Added a tiny _user() helper to keep the five templates uniform.

2. get_architecture_overview detail_level="minimal" (token-efficiency)

The default output embeds every community's full members list and every individual cross-community edge. On a 2,878-node repo (ebm-copilot — TS/Python/Bash/R, 390 files) this measured 655KB — over the MCP tool output token cap, forcing fallback-to-file and defeating the whole point of the token-efficient API.

The token-efficiency preamble already tells the LLM to call get_architecture_overview(detail_level="minimal") first, but minimal wasn't actually a supported value — so this PR finally makes the prompt's own contract work.

minimal mode:

• communities → keep only {id, name, size, cohesion, dominant_language}, drop members
• cross_community_edges → aggregate by (source_community, target_community) pair → {source_community (name), target_community (name), edge_count, top_kinds}
• warnings → unchanged

Measured on ebm-copilot:

mode	size	reduction
standard	655 KB	—
minimal	5.4 KB	120×

Output stays useful — the top pair v1-route ↔ services-list (318 CALLS edges) immediately surfaces the hottest coupling.

Test plan

• uv run pytest tests/test_prompts.py — 34 passed (updated to assert msg.role / msg.content.text)
• uv run pytest tests/test_tools.py — 3 new cases for minimal mode (drops members, aggregates pairs, summary label)
• uv run pytest tests/test_integration_v2.py — step 9 updated for Message attribute access
• Full suite: 1233 passed, 1 skipped, 2 xpassed. The 6 remaining failures are pre-existing pytest-asyncio plugin issues in tests/test_main.py::TestApplyToolFilter (the test file uses @pytest.mark.asyncio but pytest-asyncio is not a dev dep — out of scope here).
• uvx ruff check on changed files — clean.
• End-to-end: architecture_map prompt now renders through fastmcp's in-process Client (verified the exact MCP path the slash command uses).
• Smoke test on real ebm-copilot graph confirms the 120× reduction.

🤖 Generated with Claude Code