fix(jrag): honor JAVA_CODEBASE_RAG_SOURCE_ROOT over a stray ancestor marker (#387)

jrag._resolve_cfg passed source_root=discover_project_root(cwd) into
resolve_operator_config, which only honors JAVA_CODEBASE_RAG_SOURCE_ROOT when
source_root is None. So any ancestor .java-codebase-rag marker silently
overrode the env var that pipeline.subprocess_env sets for the cocoindex child
(and that operators/tests rely on), making `jrag imports/outline <rel-file>`
resolve against the wrong root → "file not found", exit 2.

Pass source_root=None instead, mirroring cli._resolved_from_ns: resolution is
now env var → YAML source_root → discover_project_root(cwd). Interactive use
(discovery fallback) is unchanged; the #383 config_source pointer is preserved.

Regression test creates a non-empty ancestor marker + sets the env var and
asserts the env var wins (fails on pre-fix code, passes with the fix).

Co-authored-by: Claude <noreply@anthropic.com>

Author: HumanBean17

java_codebase_rag/jrag.py

@@ -1096,20 +1096,26 @@ def _core_parser() -> argparse.ArgumentParser:
 def _resolve_cfg(args: argparse.Namespace):  # type: ignore[no-untyped-def]
     """Resolve operator config (reuses the operator's cocoindex-free resolver).
 
-    Same pattern as ``java_codebase_rag.cli._resolved_from_ns``: walks up from
-    cwd to find a project root (config file or ``.java-codebase-rag/`` index),
-    applies CLI ``--index-dir`` if given, and calls ``apply_to_os_environ`` so
-    downstream modules see a consistent env (critically: SBERT_MODEL for
-    ``jrag search`` in PR-JRAG-4).
+    Mirrors ``java_codebase_rag.cli._resolved_from_ns``: pass ``source_root=None``
+    so ``resolve_operator_config`` honors ``JAVA_CODEBASE_RAG_SOURCE_ROOT`` first,
+    then a YAML ``source_root`` field, then walks up from cwd to find a project
+    root. Passing a discovered root explicitly here would OVERRIDE a set env var
+    whenever any ancestor dir has a ``.java-codebase-rag`` marker — silently
+    ignoring the documented subprocess source-root mechanism that
+    ``pipeline.subprocess_env`` sets for the cocoindex child (and that operators
+    set directly).
 
     When the anchor is an index dir with no YAML beside it, resolution follows
     that index's ``config_source`` pointer (see ``config._effective_config_dir``)
     so a config living in a sibling dir is still found from inside a microservice.
+    Applies CLI ``--index-dir`` if given and calls ``apply_to_os_environ`` so
+    downstream modules see a consistent env (critically SBERT_MODEL for ``jrag
+    search``).
     """
-    from java_codebase_rag.config import discover_project_root, resolve_operator_config
+    from java_codebase_rag.config import resolve_operator_config
 
     cfg = resolve_operator_config(
-        source_root=discover_project_root(Path.cwd()),
+        source_root=None,
         cli_index_dir=getattr(args, "index_dir", None),
     )
     cfg.apply_to_os_environ()

tests/test_jrag_status.py

@@ -21,6 +21,8 @@
 import sys
 from pathlib import Path
 
+import pytest
+
 
 def _jrag_exe() -> str:
     """Locate the installed ``jrag`` entry point next to the venv interpreter."""
@@ -172,3 +174,45 @@ def test_jrag_help_lists_status_subcommand() -> None:
     assert "status" in proc.stdout
     # The --offset flag must NOT appear in the top-level help.
     assert "--offset" not in proc.stdout
+
+
+# ----- Config resolution: env var beats a stray ancestor marker -----
+
+
+def test_resolve_cfg_honors_source_root_env_over_stray_marker(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """JAVA_CODEBASE_RAG_SOURCE_ROOT must win over a stray ``.java-codebase-rag``
+    marker in an ancestor of cwd.
+
+    Regression: ``jrag._resolve_cfg`` used to pass ``discover_project_root(cwd)``
+    as the explicit ``source_root``, which silently overrode the env var whenever
+    any ancestor dir had a non-empty ``.java-codebase-rag/`` index — the
+    documented subprocess source-root mechanism ``pipeline.subprocess_env`` sets
+    for the cocoindex child (and that the traversal tests rely on). It now passes
+    ``source_root=None`` so ``resolve_operator_config`` honors the env var first
+    (mirroring ``cli._resolved_from_ns``).
+    """
+    from java_codebase_rag.jrag import _resolve_cfg
+
+    real_root = tmp_path / "real-source"
+    real_root.mkdir()
+    # cwd sits inside a dir whose PARENT has a stray non-empty .java-codebase-rag/
+    # index dir — the hijack condition that used to override the env var.
+    # (_has_index_dir requires the marker dir to be non-empty: any(iterdir()).)
+    workdir = tmp_path / "has-marker" / "sub"
+    workdir.mkdir(parents=True)
+    stray_idx = tmp_path / "has-marker" / ".java-codebase-rag"
+    stray_idx.mkdir()
+    (stray_idx / "cocoindex.db").write_bytes(b"")  # non-empty → recognized as marker
+    monkeypatch.chdir(workdir)
+    monkeypatch.setenv("JAVA_CODEBASE_RAG_SOURCE_ROOT", str(real_root))
+    monkeypatch.delenv("JAVA_CODEBASE_RAG_INDEX_DIR", raising=False)
+
+    class _Args:
+        index_dir = None
+
+    cfg = _resolve_cfg(_Args())  # type: ignore[arg-type]
+    assert cfg.source_root == real_root.resolve(), (
+        "JAVA_CODEBASE_RAG_SOURCE_ROOT must win over a stray ancestor marker"
+    )