add trace pruning collapsing cross-service (PR-TRACE-1b)

[HumanBean17]

Summary

• prune_roles soft gate: edges to pruned-role nodes are recorded in result, but the node is not added to the next BFS frontier. stats.nodes_pruned_role counts pruned nodes.
• fan_out_cap with ranking: per-node cap on candidate edges sorted by confidence (desc) → role priority (desc) → fqn (asc). Scaffolding edges (DECLARES_CLIENT, DECLARES_PRODUCER) are exempt. stats.nodes_pruned_fan_out counts dropped nodes.
• collapse_trivial post-BFS pass: degree-1 intermediates (1 in + 1 out CALLS, role OTHER) are collapsed into single edges with collapsed=True and collapsed_intermediates=[B.id]. parent_edge_id is recomputed for child edges referencing removed edges. stats.edges_collapsed_trivial counts collapsed chains.
• Cross-service boundary detection: when HTTP_CALLS/ASYNC_CALLS is in edge_types, BFS follows scaffolding edges (consuming a hop) to Client/Producer, then follows cross-service edges to downstream Route. Edges marked cross_service_boundary=True. Downstream node included in nodes dict but NOT added to frontier (boundary-stop).

Test plan

• 33 unique tests in tests/test_mcp_trace.py (31 passed, 2 skipped on fixture graph gaps)
• test_trace_filter_vs_prune_roles replaces 1a's test_trace_prune_roles_param_accepted_noop stub
• 10 new 1b tests: prune_roles, fan_out_cap, scaffolding exempt, collapse_trivial, collapse disabled, parent_edge_id consistency, cross-service HTTP, cross-service async, cross-service attrs, boundary stops
• .venv/bin/ruff check . clean
• .venv/bin/python -m pytest tests -v green (648 passed, 11 skipped)
• Sentinel check: git diff experimental -- mcp_v2.py kuzu_queries.py server.py empty

Scope

• mcp_trace.py — pruning, collapsing, cross-service implementation
• tests/test_mcp_trace.py — 11 new/replaced tests
• No changes to mcp_v2.py, kuzu_queries.py, server.py, build_ast_graph.py, java_ontology.py

🤖 Generated with Claude Code

[HumanBean17]

PR-TRACE-1b Review

Solid implementation — clean separation between the four features, good test coverage, and consistent with project conventions. A few items:

Medium

• Vacuous assertion in test_trace_cross_service_edge_attrs — len(xe.attrs) >= 0 is always true. Should assert on specific attribute keys or use > 0.

Low

• _collapse_trivial_chains is single-pass. For A→B→C→D where B and C are both trivial intermediates, only B gets collapsed (collapsed edges are skipped on line 65). Fine if intentional, but worth a docstring note about the single-pass limitation. If multi-level chains are expected, consider looping until collapsed_count == 0.

• Silent except Exception: continue in cross-service handling (~lines 279, 340). A malformed graph could silently produce incomplete results. Consider logging to stderr (safe per AGENTS.md — server is stdio-based).

• Per-node _neighbors_batched calls in cross-service (~line 312). Each Client/Producer node triggers a separate query. Could be batched if this becomes a bottleneck at scale.

• _collapse_trivial_chains mutates caller's nodes, edges, and edge_id_map in-place. A brief docstring note about mutation would help future readers.

What looks good

• prune_roles soft gate vs NodeFilter hard gate — clean semantic distinction
• Scaffolding exemption from fan_out_cap — sensible design
• parent_edge_id recomputation after collapse
• Boundary-stop semantics (node in nodes, not in frontier)
• Tests assert on invariants, not exact counts — consistent with testing philosophy
• test_trace_filter_vs_prune_roles is a good upgrade from the previous stub