Summary

Two bugs in src/mcp/mcp.c surfaced during a real-world MCP tool sweep against a mode=full index of a 12,770-node Python repo on macOS arm64. Both have minimal, targeted fixes.

Bug 1 — list_projects returns {"projects":[]} even when projects exist

Symptom: After index_repository(...) succeeds, the project is fully queryable via index_status, search_code, search_graph, query_graph, get_code_snippet, trace_path, etc., but list_projects returns an empty array.

Root cause: is_project_db_file() (mcp.c:846, also inlined inside collect_db_project_names at mcp.c:791) excludes any filename with a tmp- prefix. But cbm_project_name_from_path() legitimately produces tmp-... project names for source roots under /tmp/ — verified by an existing in-tree test:

// tests/test_pipeline.c:1903
{"/tmp/bench/erlang/lib/stdlib/src", "tmp-bench-erlang-lib-stdlib-src"},

Any project rooted under /tmp was silently filtered from list_projects + the available_projects hint inside build_project_list_error.

Fix: drop the tmp- prefix exclusion entirely. SQLite's actual journal/WAL sidecar files end with -journal, -wal, -shm (not .db) and never collide with the .db suffix is_project_db_file already requires, so removing the prefix filter is safe. While here, refactor collect_db_project_names() to call is_project_db_file() so the filter lives in one place.

Bug 2 — trace_path returns function not found for fully-qualified names

Symptom: Calling trace_path with a fully-qualified function name ("<project>.module.submodule.func") returns {"error":"function not found", ...} even when the node exists. The tool's own error hint instructs callers to "fully qualify" the name, but doing so causes the lookup to miss every time — the documented "fully qualify for precise match" UX inverts what actually works.

Root cause: handle_trace_call_path() (mcp.c:1887) only calls cbm_store_find_nodes_by_name(), which queries WHERE project = ?1 AND name = ?2 against the bare name column. The qualified name is never tried.

Fix: when the bare-name lookup returns 0 nodes, fall back to cbm_store_find_node_by_qn() (already exists in store.c, used elsewhere). If that resolves, wrap the single result into the same nodes[] array the downstream BFS expects, so the rest of the function is unchanged.

Verification

Built locally with make -f Makefile.cbm cbm, replaced the cached binary at ~/Library/Caches/codebase-memory-mcp/0.6.0/codebase-memory-mcp on macOS arm64, restarted the host service, and verified against a real Users-jwiley-dev-orca index (12,770 nodes / 36,859 edges) plus a synthetic tmp-cbm_sandbox project rooted at /tmp/cbm_sandbox:

• Pre-patch: list_projects returned {"projects":[]}.
  Post-patch: returns both tmp-cbm_sandbox and Users-jwiley-dev-orca.

• Pre-patch: trace_path(function_name="Users-jwiley-dev-orca.src.services.session_enricher.enrichment_service.run_enrichment") returned function not found.
  Post-patch: returns identical callees (29 nodes across hops 1+2) and callers (2 nodes) to the bare-name lookup trace_path(function_name="run_enrichment").

Existing tests untouched.

Test plan

• Verified patched binary builds cleanly with -O2 -Wall -Wextra -Werror on macOS arm64
• Verified list_projects now surfaces a tmp-cbm_sandbox project (was previously filtered)
• Verified trace_path with bare name (control) still works
• Verified trace_path with fully-qualified name returns identical traversal
• CI on this PR (Linux + Windows builds + make -f Makefile.cbm test)