test: session graph fixtures and builders (PR-1 fast loop)

[HumanBean17]

Scope

Implements PR-1 (fixture fast loop) from plans/PLAN-TEST-SUITE-FAST-LOOP.md: session-scoped Kuzu builds per static corpus, shared tests/_builders.py helpers, Tier 1/2 test migrations, Tier 3 per-test builds delegating to _builders, and tests/README.md updates (merge gate intent, three tiers, consumer matrix, iteration/skills pointers).

Out of this PR: GitHub Actions workflow and branch protection (PR-3). No production code. No ontology bump / no re-index.

Highlights

• tests/_builders.py: build_graph_tables_to, build_kuzu_to, and build_kuzu_*_into wrappers (thin façade over production passes).
• tests/conftest.py: Bank session graph pass1–5 + write, no pass6; session DB paths for call_graph_smoke, route_extraction_smoke, cross_service_smoke, fqn_collision_smoke, http_caller_smoke; graph_tables_cross_service_smoke for read-only in-memory tests; TYPE_CHECKING for GraphTables on the session tables fixture.
• Singleton safety: kuzu_graph_route_extraction_smoke and call-graph roundtrip tests use KuzuGraph(path) where appropriate instead of fighting KuzuGraph.get.
• test_mcp_v2.test_find_client_by_target_service: Seeds from list_clients() rows with real target_service (display fqn first token was wrong once pass5 added more clients). Documented in README as a test fix, not fixture scope creep.
• Tier-3 mutable copies: test_cross_service_resolution_flag.py and test_client_role_rename.py still copy/mutate cross_service_smoke under tmp_path; they now call _builders instead of inlining passes.

Consumer matrix

See tests/README.md (section CI merge gate and fixture tiers) for the bank-chat / test_call_invariant rows and pass-depth rationale.

Validation

.venv/bin/ruff check tests/
JAVA_CODEBASE_RAG_RUN_HEAVY=0 .venv/bin/python -m pytest \
  tests/test_call_graph_smoke_roundtrip.py tests/test_ast_graph_build.py tests/test_kuzu_queries.py \
  tests/test_call_edges_e2e.py tests/test_call_invariant.py tests/test_feign_not_exposer.py \
  tests/test_call_edge_matching.py tests/test_client_hint_recovery.py tests/test_brownfield_routes.py \
  tests/test_brownfield_clients.py tests/test_client_node_extraction.py tests/test_assign_endpoint_client_extraction.py \
  tests/test_mcp_v2.py tests/test_call_graph_receiver_resolution.py tests/test_cross_service_resolution_flag.py \
  tests/test_client_role_rename.py -q

Result: 188 passed, 3 skipped (expected skips in test_mcp_v2). Full pytest tests should still be run before merge once CI exists (PR-3) or locally for confidence.

Timing

Before/after wall-time for a full pytest tests run was not captured in this environment; worth a quick note in a follow-up comment after you run it locally.