fix(graph): incremental global pass5/6 runs pass4; preserve stub callee role (#352)

[HumanBean17]

What

increment produced a graph that differed from a clean full rebuild in two ways, both in the incremental global pass5/6 step (#352).

Divergence #1 — missing pass4 → HTTP_CALLS match drift

The global pass5/6 step rebuilt tables_for_global via pass1 → pass5 → pass6 with no pass4 (routes/EXPOSES). So tables_for_global.exposes_rows / routes_rows stayed empty; pass5 links Feign HTTP_CALLS to routes via exposes_rows and pass6 matches against routes_rows, so the HTTP_CALLS match outcome drifted from the full rebuild (e.g. on bank-chat /chat/joinOperator Feign edges flipped unresolved → phantom).

Fix: run pass4_routes on tables_for_global between pass1 and pass5 — exactly what main() does for a full rebuild.

Divergence #2 — callee_declaring_role collapsed for out-of-scope stubs

_load_existing_types loaded out-of-scope types as annotation-less TypeDecl stubs without their persisted role, so _write_nodes_impl recomputed role via resolve_role_and_capabilities on the empty stub and collapsed CONTROLLER/SERVICE/… to the default OTHER. That wrong role was stored in type_role_by_node_id[stub_id], which feeds _callee_declaring_role_at_write → every CALLS edge into the type got the wrong callee_declaring_role after any increment.

Fix: _load_existing_types now loads s.role and seeds type_role_by_node_id for stubs; _write_nodes_impl trusts the persisted role for loaded_from_db stubs instead of overwriting it (the staged stub row is already filtered out of the write via stub_ids, so this only corrects the in-memory role map).

Test

Strengthened test_incremental_bulk_write_equivalent_to_full_rebuild:

• _graph_state now also dumps the HTTP_CALLS / ASYNC_CALLS match histograms and the CALLS callee_declaring_role multiset (previously only node count, edge counts, and type roles — so both divergences shipped green).
• New corpus: a @RestController Cal (route + CONTROLLER callee, left out of scope), a @FeignClient, and a Caller that is touched.

Divergence #2 is directly tested (TDD red → green): before the fix, incremental gave {'OTHER': 2} vs full {'CONTROLLER': 2} for callee_declaring_role; after, they match.

Divergence #1 is fixed by mirroring main() exactly. The histogram-equality assertion is a non-trivial regression guard (the corpus has a real HTTP_CALLS edge), but this minimal corpus's Feign outcome is phantom in both paths — the unresolved→phantom drift in the issue depends on bank-chat's microservice-configured Feign match, which isn't reproducible inline. The fix is the canonical correct behavior (run pass4 on the global tables).

.venv/bin/python -m pytest tests/test_incremental_graph.py -q      # 33 passed
.venv/bin/python -m pytest tests/test_mcp_v2.py -q                 # 107 passed, 1 skipped
.venv/bin/python build_ast_graph.py --source-root tests/bank-chat-system --ladybug-path /tmp/cg.lbug   # exit 0

Notes

• No re-index required and no ontology_version bump — full-rebuild semantics are unchanged; this makes increment match the canonical full index. No env-var change.

Closes #352.

🤖 Generated with Claude Code

[HumanBean17]

Addressed a review finding: the http_calls_match histogram couldn't trip on divergence #1 — the corpus has no microservice, so the Feign edge resolves to 'phantom' in both the incremental and full paths, making that assertion non-load-bearing despite its message.

• Added test_incremental_global_pass_runs_pass4_on_global_tables: a spy on pass4_routes asserting it runs on a DISTINCT global tables instance (the full-source rebuild) alongside the changed-files tables. Removing the global pass4 call drops the count to 1 → failure. Verified red→green.
• Corrected the equivalence test's http_calls_match message to state plainly it's a consistency guard pointing at the structural test for divergence AST by Opus #1.