fix(cache): recover semantic cache groups with malformed source_file by basename

CHANGELOG.md

@@ -4,6 +4,7 @@ Full release notes with details on each version: [GitHub Releases](https://githu
 
 ## 0.9.61 (unreleased)
 
+- Fix: `save_semantic_cache` now recovers a group whose reported `source_file` never resolves to a real file, by an unambiguous basename match against the already dispatched allowlist, instead of silently dropping it — an incremental run driven by a weak or local backend no longer redoes already processed chunks forever when the adaptive-retry split path loses track of a source file (#2973, thanks @ayushcodes10).
 - Fix: `graphify.serve` now imports cleanly on Python 3.12 and 3.13. The `chinese` extra pins `jieba-py` from 3.12 onward (0.9.60 mistakenly kept the old `jieba` until 3.14, and its invalid regex escapes are a hard error on 3.12+), and the jieba import now suppresses the tokenizer's `SyntaxWarning` regardless of message or line so it never escalates under `-W error`.
 - Fix: the git hook's rebuild-root guard now rejects a symlink-loop or dangling `.graphify_root` on Python 3.13, whose `Path.resolve()` no longer raises on a loop — the saved root must resolve to a real directory inside the repo before it is adopted.
 

graphify/cache.py

@@ -1490,6 +1490,31 @@ def resolved_source_path(value: str | Path) -> Path:
     if allowed_source_files is not None:
         allowed_paths = {source_path(path) for path in allowed_source_files}
 
+    def _recover_group_path(fpath: str) -> tuple[Path, Path]:
+        """Return ``(cache_path, resolved_path)`` for one ``by_file`` group,
+        recovering an unresolvable ``source_file`` via an unambiguous
+        basename match against ``allowed_paths`` when one is available (#2973).
+
+        The adaptive-retry split path (``llm.py``'s bisect-and-retry on a
+        chunk that overflowed a weak/local backend's context) sometimes
+        re-prompts with a reduced file subset and loses track of which of
+        the original chunk's files a given node came from, so its
+        ``source_file`` never resolves to a real path at all. Recovering it
+        against the known-good, already-dispatched allowlist -- and ONLY
+        when the basename is unambiguous there -- lets that group's nodes
+        and edges reach the cache instead of silently vanishing on every
+        incremental run. A genuinely bogus or ambiguous basename still
+        falls through unrecovered to the existing skip behavior below.
+        """
+        cache_path = source_path(fpath)
+        resolved = resolved_source_path(fpath)
+        if not resolved.is_file() and allowed_paths is not None:
+            candidates = [ap for ap in allowed_paths if ap.name == cache_path.name]
+            if len(candidates) == 1:
+                cache_path = candidates[0]
+                resolved = candidates[0]
+        return cache_path, resolved
+
     partial_paths = None
     if partial_source_files is not None:
         partial_paths = {source_path(path) for path in partial_source_files}
@@ -1506,9 +1531,9 @@ def resolved_source_path(value: str | Path) -> Path:
 
     def group_skipped(fpath: str) -> bool:
         """Mirror the write-loop skip condition for one source_file group."""
-        p = resolved_source_path(fpath)
+        cache_path, p = _recover_group_path(fpath)
         return not p.is_file() or (
-            allowed_paths is not None and source_path(fpath) not in allowed_paths
+            allowed_paths is not None and cache_path not in allowed_paths
         )
 
     # Dangling-reference pruning (#1916). A node group is skipped by the write
@@ -1565,9 +1590,26 @@ def hyperedge_dangles(h: dict) -> bool:
     saved = 0
     skipped_not_file = 0
     for fpath, result in by_file.items():
-        cache_path = source_path(fpath)
-        p = resolved_source_path(fpath)
+        cache_path, p = _recover_group_path(fpath)
         if p.is_file():
+            if cache_path != source_path(fpath):
+                # #2973: recovery only redirected the WRITE KEY. Each item in
+                # this group still carries the original unresolvable
+                # source_file string, and _semantic_entry_matches_path
+                # rejects an entry on read if any item's source_file doesn't
+                # match the path it was loaded under -- so leaving the old
+                # value in place would write a "successful" entry that can
+                # never actually be read back, silently reproducing the same
+                # loss this recovery exists to fix.
+                corrected = _normalize_value(str(cache_path))
+                result = {
+                    **result,
+                    "nodes": [{**n, "source_file": corrected} for n in result["nodes"]],
+                    "edges": [{**e, "source_file": corrected} for e in result["edges"]],
+                    "hyperedges": [
+                        {**h, "source_file": corrected} for h in result["hyperedges"]
+                    ],
+                }
             if allowed_paths is not None and cache_path not in allowed_paths:
                 warnings.warn(
                     "semantic cache skipped out-of-scope source_file "

tests/test_semantic_cache_basename_recovery.py

@@ -0,0 +1,91 @@
+"""#2973 — save_semantic_cache must recover a group whose reported
+source_file never resolves to a real file, by an unambiguous basename match
+against the already dispatched allowlist, instead of silently dropping it.
+
+A weak or local backend's adaptive-retry split path (llm.py bisecting a
+chunk that overflowed context and retrying) sometimes re-prompts with a
+reduced file subset and loses track of which of the original chunk's files
+a given node came from, so its source_file drifts to something that never
+resolves at all. Without recovery this silently discarded the group's nodes
+and edges from the cache on every incremental run.
+"""
+from __future__ import annotations
+
+import pytest
+
+from graphify.cache import load_cached, save_semantic_cache
+
+
+def test_malformed_but_basename_unique_path_recovers(tmp_path):
+    real = tmp_path / "sub" / "weird_named_file.py"
+    real.parent.mkdir(parents=True)
+    real.write_text("def f(): pass\n")
+
+    nodes = [{"id": "n1", "label": "f", "source_file": "lost_dir/weird_named_file.py"}]
+    saved = save_semantic_cache(nodes, [], root=tmp_path, allowed_source_files=[real])
+    assert saved == 1
+
+    cached = load_cached(real, root=tmp_path, kind="semantic")
+    assert cached is not None
+    assert {n["id"] for n in cached["nodes"]} == {"n1"}
+
+
+def test_recovered_group_edges_are_not_pruned_as_dangling(tmp_path):
+    # group_skipped (used by the dangling-reference pruning pass) and the
+    # write loop must agree a recovered group is WRITTEN, not skipped --
+    # otherwise an edge between two nodes in that same recovered group would
+    # be wrongly pruned as referencing a "skipped" id.
+    real = tmp_path / "sub" / "weird_named_file.py"
+    real.parent.mkdir(parents=True)
+    real.write_text("def f(): pass\ndef g(): pass\n")
+
+    nodes = [
+        {"id": "n1", "label": "f", "source_file": "lost_dir/weird_named_file.py"},
+        {"id": "n2", "label": "g", "source_file": "lost_dir/weird_named_file.py"},
+    ]
+    edges = [
+        {"source": "n1", "target": "n2", "relation": "calls",
+         "source_file": "lost_dir/weird_named_file.py"},
+    ]
+    saved = save_semantic_cache(nodes, edges, root=tmp_path, allowed_source_files=[real])
+    assert saved == 1
+
+    cached = load_cached(real, root=tmp_path, kind="semantic")
+    assert cached is not None
+    assert len(cached["edges"]) == 1
+
+
+def test_ambiguous_basename_stays_skipped(tmp_path):
+    a = tmp_path / "pkg_a" / "shared.py"
+    b = tmp_path / "pkg_b" / "shared.py"
+    a.parent.mkdir(parents=True)
+    b.parent.mkdir(parents=True)
+    a.write_text("def f(): pass\n")
+    b.write_text("def g(): pass\n")
+
+    nodes = [{"id": "n1", "label": "f", "source_file": "lost_dir/shared.py"}]
+    with pytest.warns(RuntimeWarning, match="do not resolve to real files"):
+        saved = save_semantic_cache(nodes, [], root=tmp_path, allowed_source_files=[a, b])
+    assert saved == 0
+    assert load_cached(a, root=tmp_path, kind="semantic") is None
+    assert load_cached(b, root=tmp_path, kind="semantic") is None
+
+
+def test_unscoped_call_with_no_allowlist_is_unaffected(tmp_path):
+    # No allowed_source_files at all: recovery must never run, so a
+    # genuinely bogus path is skipped exactly as before this fix, and a
+    # normal well formed path still resolves and saves.
+    real = tmp_path / "sub" / "weird_named_file.py"
+    real.parent.mkdir(parents=True)
+    real.write_text("def f(): pass\n")
+
+    nodes = [
+        {"id": "n1", "label": "f", "source_file": "sub/weird_named_file.py"},
+        {"id": "n2", "label": "g", "source_file": "totally/does/not/exist.py"},
+    ]
+    saved = save_semantic_cache(nodes, [], root=tmp_path)
+    assert saved == 1
+
+    cached = load_cached(real, root=tmp_path, kind="semantic")
+    assert cached is not None
+    assert {n["id"] for n in cached["nodes"]} == {"n1"}