chore: polish bundle — correctness/robustness nits (#359)

[HumanBean17]

What

A bundle of low-severity correctness/robustness nits from #359. Each independently small.

1. exclude_roles now disables role weighting (search_lancedb.py) — skip_role_weight = bool(role or role_in), so an exclude_roles-only filter still applied role weights (asymmetric with role/role_in). Now includes exclude_roles.
2. _parse_ladybug_json quotes only key positions (ladybug_queries.py) — the old (\w+): regex matched word: inside values too (e.g. a URL https://...), corrupting them. Now quotes keys only after { / , / [. Also fixed the misleading "counts_json" log label (the helper parses many graph_meta blobs). New unit test test_parse_ladybug_json_handles_colon_in_values.
3. resolve rejects wildcards consistently (mcp_v2.py) — * / ? silently returned status='none', while search/find/neighbors reject them. Now success=False with a pointer to search(query=...). Renamed/updated the wildcard test.
4. Client-kind literals de-leaked (java_ontology/ast_java/build_ast_graph) — feign_method/rest_template/web_client were emitted and matched directly. Added named CLIENT_KIND_* constants (exported in __all__) and reference them at every emit/match site so a rename in VALID_CLIENT_KINDS can't silently desync callers.
5. CALLS dedup key includes call_site_byte (build_ast_graph.py) — two call sites of the same method on the same source line (same arg_count) no longer collapse into one edge. Regenerated the bank-chat baseline CALLS count (678 → 684; sampled CALLS first-3 are byte-identical since the new edges land later in insertion order) and mirrored the byte-inclusive key in the dedup test.
6. Malformed YAML no longer silently swallowed (config.py) — load_yaml_mapping caught bare Exception and returned {}. Now catches yaml.YAMLError and emits a stderr hint.
7. table_names() → list_tables() (config.py, cli.py) — the lancedb DeprecationWarning is gone (lance_optimize already preferred list_tables with a fallback).

Deferred (with justification)

• FastMCP.run_stdio_async "coroutine never awaited" RuntimeWarning (server.py:752): benign — the server runs fine; run_stdio_async is a clean coroutine properly awaited by asyncio.run, and the root cause appears to be inside the FastMCP stdio transport on premature stdin close. A speculative entry-point change risks destabilizing MCP server startup for no functional gain, so I left it rather than guess.

Test

.venv/bin/python -m pytest tests/test_ast_graph_build.py tests/test_mcp_v2.py tests/test_incremental_graph.py \
  tests/test_ladybug_queries.py tests/test_call_invariant.py tests/test_call_graph_smoke_roundtrip.py \
  tests/test_call_edges_e2e.py tests/test_bank_chat_brownfield_integration.py tests/test_brownfield_clients.py \
  tests/test_brownfield_routes.py tests/test_config.py tests/test_search_lancedb.py tests/test_client_hint_recovery.py -q
# 109 + 30 + ... all green; no regressions
.venv/bin/python -m pytest tests/test_java_codebase_rag_cli.py -q   # 61 passed (table_names deprecation warnings gone)

The bank-chat full-build smoke also passes (build_ast_graph.py --source-root tests/bank-chat-system).

Notes

• The CALLS dedup change (PR-A1: Route schema + literal extractor (B2a) #5) is the only one that changes indexed output: bank-chat gains 6 CALLS edges that were previously wrongly collapsed. The baseline regeneration reflects this intended correction. No ontology_version bump (full-rebuild semantics are the canonical reference; this makes CALLS edges more granular, not differently typed).

Closes #359.

🤖 Generated with Claude Code

[HumanBean17]

Addressed 3 findings from a multi-agent review pass:

1. counts.calls undercount (medium) — _write_meta deduped CALLS with the pre-widening 4-tuple key while _write_edges wrote with the 5-tuple (+ call_site_byte), so describe reported calls: 678 while the graph held 684 edges. Mirrored the 5-tuple in _write_meta and bumped the baseline counts_json.calls → 684. Verified red→green (baseline test fails with only the fixture bumped).
2. Missed client-kind literal (low) — graph_enrich.py still emitted a bare "rest_template" default at the synthesized-hint site; now uses CLIENT_KIND_REST_TEMPLATE.
3. load_yaml_mapping except narrowed too far (low) — OSError (chmod 000, stat/read TOCTOU) and UnicodeDecodeError (non-UTF-8) propagated instead of degrading to {}. Broadened to (yaml.YAMLError, OSError, UnicodeDecodeError); kept the stderr hint.

3 commits on this branch. ruff check clean; test_ast_graph_build.py + test_config.py green.