fix(jrag): honor JAVA_CODEBASE_RAG_SOURCE_ROOT over a stray ancestor marker

[HumanBean17]

Summary

Fixes the 3 test_jrag_traversal_compose.py failures (test_outline_lists_file_symbols, test_imports_resolves_graph_nodes, test_imports_text_mode_marks_unresolved) by making jrag honor JAVA_CODEBASE_RAG_SOURCE_ROOT over a stray ancestor .java-codebase-rag marker — matching resolve_operator_config's intended precedence (env var > YAML > discovery) and the java-codebase-rag CLI.

Root cause

jrag._resolve_cfg passed source_root=discover_project_root(Path.cwd()) into resolve_operator_config. That function only honors JAVA_CODEBASE_RAG_SOURCE_ROOT when source_root is None (config.py: the env-var branch is the else of if source_root is not None). So whenever any ancestor of cwd had a .java-codebase-rag/ marker, the discovered root was passed explicitly and the env var was silently ignored — even though pipeline.subprocess_env sets exactly that env var for the cocoindex child and the operator/test contract relies on it.

Symptom: jrag imports <rel-file> resolved <file> against the wrong source root and returned file not found ... (looked at <source_root>/...), exit 2.

This is the same class of hijack issue #357 addressed for $HOME, generalized: any ancestor marker beats an explicit env var. CI stayed green only because a clean checkout has no stray marker; the failures reproduced anywhere a .java-codebase-rag/ index existed up-tree from the run dir.

Fix (java_codebase_rag/jrag.py)

_resolve_cfg now passes source_root=None, mirroring cli._resolved_from_ns (which passes None unless --source-root is given). resolve_operator_config then resolves: env var → YAML source_root → discover_project_root(cwd). Interactive use (no env var) is unchanged — discovery still runs as the fallback. The config_source pointer behavior from #383 is preserved (it's part of resolve_operator_config's index-dir path, which still runs).

Test (tests/test_jrag_status.py)

test_resolve_cfg_honors_source_root_env_over_stray_marker — creates a non-empty .java-codebase-rag/ marker in an ancestor of cwd, sets JAVA_CODEBASE_RAG_SOURCE_ROOT to a different dir, and asserts _resolve_cfg resolves to the env-var root. Verified to fail on the pre-fix code and pass with the fix, so CI catches future regressions (the traversal tests only caught it when a marker happened to exist locally).

Verification

• All 3 previously-failing tests now pass (with a stray marker present — the hijack condition).
• Full jrag suite green.
• Regression test fails without the fix, passes with it.

🤖 Generated with Claude Code