diff --git a/graphify/cache.py b/graphify/cache.py
index 7122b5c58..c5edc0839 100644
--- a/graphify/cache.py
+++ b/graphify/cache.py
@@ -465,7 +465,8 @@ def cache_dir(root: Path = Path("."), kind: str = "ast",
 def load_cached(path: Path, root: Path = Path("."), kind: str = "ast",
                 cache_root: Path | None = None, prompt: "str | Path | None" = None,
                 prompt_file: "str | Path | None" = None,
-                allow_legacy: bool = True) -> dict | None:
+                allow_legacy: bool = True,
+                allow_partial: bool = False) -> dict | None:
     """Return cached extraction for this file if hash matches, else None.
 
     Cache key: SHA256 of file contents.
@@ -514,6 +515,17 @@ def load_cached(path: Path, root: Path = Path("."), kind: str = "ast",
             result = json.loads(entry.read_text(encoding="utf-8"))
         except (json.JSONDecodeError, OSError):
             return None
+        # A ``partial`` entry was produced from a truncated LLM response and
+        # covers only part of the file's symbols. Serving it as authoritative
+        # would return the incomplete node set forever until the file is
+        # re-extracted. Treat it as a cache MISS (the normal read path) so the
+        # file is re-dispatched and retried. Self-heals: a later complete
+        # extraction overwrites the same content-hash key with a non-partial
+        # entry. ``allow_partial`` is the one exception — the merge_existing
+        # checkpoint peeks at a partial prev so it can accumulate a file's slices
+        # across chunks without losing the truncated one (it stays partial).
+        if not allow_partial and isinstance(result, dict) and result.get("partial"):
+            return None
         if legacy_hit:
             _legacy_semantic_hits += 1
         # Re-anchor relative source_file fields so callers see the same
@@ -748,6 +760,23 @@ def check_semantic_cache(
     return cached_nodes, cached_edges, cached_hyperedges, uncached
 
 
+def _group_has_partial_marker(group: dict) -> bool:
+    """True if any node/edge/hyperedge in a per-file group carries the internal
+    ``_partial`` truncation marker set by the adaptive-retry give-up sites.
+
+    The marker rides the item dicts up through every chunk merge, so it reaches
+    ``save_semantic_cache`` on BOTH the incremental checkpoint path (llm.py) and
+    the final authoritative save (cli.py) without either caller having to thread
+    an extra argument — the final save would otherwise overwrite a checkpoint's
+    ``partial`` flag with a clean-looking entry.
+    """
+    for bucket in ("nodes", "edges", "hyperedges"):
+        for item in group.get(bucket, []):
+            if isinstance(item, dict) and item.get("_partial"):
+                return True
+    return False
+
+
 def save_semantic_cache(
     nodes: list[dict],
     edges: list[dict],
@@ -758,6 +787,7 @@ def save_semantic_cache(
     mode: str | None = None,
     prompt: "str | Path | None" = None,
     prompt_file: "str | Path | None" = None,
+    partial_source_files: Iterable[str | Path] | None = None,
 ) -> int:
     """Save semantic extraction results to cache, keyed by source_file.
 
@@ -782,6 +812,13 @@ def save_semantic_cache(
     file, but a model must not be able to replace that file's complete cache
     entry unless the file was part of the current extraction batch (#1757).
 
+    When ``partial_source_files`` is provided, entries for those files are
+    stamped ``partial: True`` — the extraction was truncated, so the entry is
+    incomplete and :func:`load_cached` must treat it as a miss. Partial-ness is
+    ALSO detected intrinsically from a ``_partial`` marker on any grouped item,
+    so the flag survives even when a caller (e.g. cli.py's final save) does not
+    pass ``partial_source_files``.
+
     ``prompt`` is the extraction prompt that produced these results — text, or
     a Path to the prompt file. It stamps entries into the p{fingerprint}/
     namespace so a later run under a different prompt re-extracts rather than
@@ -824,6 +861,10 @@ def save_semantic_cache(
     if allowed_source_files is not None:
         allowed_paths = {resolved_source_path(path) for path in allowed_source_files}
 
+    partial_paths = None
+    if partial_source_files is not None:
+        partial_paths = {resolved_source_path(path) for path in partial_source_files}
+
     def group_skipped(fpath: str) -> bool:
         """Mirror the write-loop skip condition for one source_file group."""
         p = resolved_source_path(fpath)
@@ -898,14 +939,32 @@ def save_semantic_cache(
                 # then stamp the result as current-vintage — the exact mixing
                 # #1939 is about, made unfixable because the entry now claims a
                 # prompt that only produced half of it.
+                # allow_partial=True: a file split into slices across chunks
+                # accumulates here; if an earlier slice truncated, keep its nodes
+                # in the union AND let the entry stay partial (the _partial
+                # markers ride through, so is_partial below re-detects it) rather
+                # than a later clean slice silently replacing it and promoting the
+                # half-file to complete.
                 prev = load_cached(p, root, kind=kind, prompt=prompt,
-                                   prompt_file=prompt_file, allow_legacy=False)
+                                   prompt_file=prompt_file, allow_legacy=False,
+                                   allow_partial=True)
                 if prev:
                     result = {
                         "nodes": (prev.get("nodes", []) or []) + result["nodes"],
                         "edges": (prev.get("edges", []) or []) + result["edges"],
                         "hyperedges": (prev.get("hyperedges", []) or []) + result["hyperedges"],
                     }
+            # A file is partial if the caller named it OR any of its grouped
+            # items carries the intrinsic ``_partial`` marker. Stamp the flag
+            # load_cached keys off; copy so the caller's dict is never mutated.
+            # A later complete extraction overwrites this same content-hash key
+            # with a non-partial entry that then serves normally.
+            is_partial = (
+                (partial_paths is not None and p in partial_paths)
+                or _group_has_partial_marker(result)
+            )
+            if is_partial:
+                result = {**result, "partial": True}
             save_cached(p, result, root, kind=kind, prompt=prompt, prompt_file=prompt_file)
             saved += 1
     return saved
diff --git a/graphify/cli.py b/graphify/cli.py
index abf278378..2c18e2e2e 100644
--- a/graphify/cli.py
+++ b/graphify/cli.py
@@ -58,12 +58,19 @@ def _stamped_manifest_files(
     files_by_type: dict[str, list[str]],
     sem_result: dict,
     root: Path,
+    partial_source_files: "set[str] | None" = None,
 ) -> dict[str, list[str]]:
     """Manifest-safe files dict: only stamp semantic files that actually
     produced output (cache hit or fresh extraction). Files whose chunk failed
     have no source_file entry in sem_result — leaving their semantic_hash
     empty so detect_incremental re-queues them (#933).
 
+    A file in ``partial_source_files`` DID produce output this run, but only a
+    truncated fragment of it, so it is excluded from stamping too — otherwise
+    detect_incremental would see it "done" and never re-dispatch it, leaving the
+    incomplete node set live forever on the warm-incremental path. Same #933
+    mechanism: leave it unstamped and it is re-queued next run.
+
     Both sides of the membership test are resolved against the scan ``root``
     before comparing (#1897): node/edge/hyperedge ``source_file`` values are
     root-relative on a fresh extraction while ``files_by_type`` entries are
@@ -95,11 +102,13 @@ def _stamped_manifest_files(
             sf = item.get("source_file", "")
             if sf:
                 sem_extracted.add(_resolve(sf))
+    partial_resolved = {_resolve(p) for p in (partial_source_files or set())}
     sem_types = {"document", "paper", "image"}
     return {
         ftype: [
             f for f in flist
-            if ftype not in sem_types or _resolve(f) in sem_extracted
+            if ftype not in sem_types
+            or (_resolve(f) in sem_extracted and _resolve(f) not in partial_resolved)
         ]
         for ftype, flist in files_by_type.items()
     }
@@ -2568,6 +2577,11 @@ def dispatch_command(cmd: str) -> None:
             "nodes": [], "edges": [], "hyperedges": [],
             "input_tokens": 0, "output_tokens": 0,
         }
+        # Semantic files whose extraction truncated this run. They are left
+        # unstamped in the manifest so detect_incremental re-queues them next run
+        # (mirrors the #933 failed-chunk handling); captured below before the
+        # _partial markers are stripped from the corpus.
+        _partial_semantic_files: set[str] = set()
         sem_cache_hits = 0
         sem_cache_misses = 0
         # Deep mode uses its own namespace (cache/semantic-deep/) so deep and
@@ -2673,6 +2687,17 @@ def dispatch_command(cmd: str) -> None:
                     )
                 except Exception as exc:
                     print(f"[graphify extract] warning: could not write semantic cache: {exc}", file=sys.stderr)
+                # Record which files truncated (before the markers are stripped)
+                # so they are left unstamped in the manifest and re-queued on the
+                # next incremental run. The save above consumed the marker to
+                # stamp affected cache entries partial: True; strip it before the
+                # corpus feeds the graph so it never leaks into graph.json.
+                from graphify.llm import (
+                    _partial_source_files as _partial_sf,
+                    _strip_partial_markers as _strip_partial,
+                )
+                _partial_semantic_files = set(_partial_sf(fresh))
+                _strip_partial(fresh)
                 sem_result["nodes"].extend(fresh.get("nodes", []))
                 sem_result["edges"].extend(fresh.get("edges", []))
                 sem_result["hyperedges"].extend(fresh.get("hyperedges", []))
@@ -2752,7 +2777,8 @@ def dispatch_command(cmd: str) -> None:
         # Path normalization against the scan root happens inside the helper
         # (#1897) so fresh root-relative source_files match detect()'s
         # absolute file lists.
-        _manifest_files = _stamped_manifest_files(files_by_type, sem_result, target)
+        _manifest_files = _stamped_manifest_files(files_by_type, sem_result, target,
+                                                   partial_source_files=_partial_semantic_files)
 
         # Full-scan manifest saves prune rows for in-root files that left the
         # scan corpus but still exist on disk (#1908). The corpus must be the
diff --git a/graphify/llm.py b/graphify/llm.py
index 3ce10f283..3cbaf0eaa 100644
--- a/graphify/llm.py
+++ b/graphify/llm.py
@@ -1636,6 +1636,50 @@ def _looks_like_context_exceeded(exc: BaseException) -> bool:
     return any(marker in msg for marker in _CONTEXT_EXCEEDED_MARKERS)
 
 
+def _mark_partial(result: dict) -> None:
+    """Tag every node/edge/hyperedge in a truncated chunk result with an internal
+    ``_partial`` marker.
+
+    A chunk whose LLM response was truncated (`finish_reason="length"`) and could
+    not be recovered by splitting yields a PARTIAL node set. Left unmarked, that
+    set is checkpointed and (via the final save) written to the content-hash
+    semantic cache as authoritative, so it is served forever until the file
+    content changes or ``--force``. The marker rides these item dicts up through
+    every chunk merge (which concatenate the same object references) so it reaches
+    ``save_semantic_cache`` on both the checkpoint and the final-save paths, which
+    stamp the entry ``partial: True``; ``load_cached`` then treats it as a miss.
+    """
+    for bucket in ("nodes", "edges", "hyperedges"):
+        for item in result.get(bucket, []):
+            if isinstance(item, dict):
+                item["_partial"] = True
+
+
+def _partial_source_files(result: dict) -> list[str]:
+    """Source files that carry at least one ``_partial`` item in ``result``."""
+    seen: set[str] = set()
+    for bucket in ("nodes", "edges", "hyperedges"):
+        for item in result.get(bucket, []):
+            if isinstance(item, dict) and item.get("_partial"):
+                sf = item.get("source_file")
+                if sf:
+                    seen.add(str(sf))
+    return sorted(seen)
+
+
+def _strip_partial_markers(result: dict) -> None:
+    """Remove the internal ``_partial`` marker from every item in ``result``.
+
+    Call this only AFTER the semantic cache has been saved (the save consumes the
+    marker to stamp affected entries ``partial: True``). Stripping it keeps the
+    internal flag out of the graph.json nodes/edges the corpus result feeds into.
+    """
+    for bucket in ("nodes", "edges", "hyperedges"):
+        for item in result.get(bucket, []):
+            if isinstance(item, dict):
+                item.pop("_partial", None)
+
+
 def _extract_with_adaptive_retry(
     chunk: list[Path],
     backend: str,
@@ -1769,17 +1813,25 @@ def _extract_with_adaptive_retry(
             return _merge_two([halves[0]], [halves[1]])
         print(
             f"[graphify] single-file chunk {unit_path(chunk[0])} truncated at "
-            f"max_completion_tokens — partial result kept",
+            f"max_completion_tokens — partial result kept (not cached as complete)",
             file=sys.stderr,
         )
+        # The node set is incomplete; mark it so it is not promoted to the
+        # semantic cache as authoritative and is re-dispatched next run.
+        _mark_partial(result)
         return result
 
     if _depth >= max_depth:
         print(
             f"[graphify] chunk of {len(chunk)} still truncated at recursion "
-            f"depth {_depth} (max {max_depth}) — partial result kept",
+            f"depth {_depth} (max {max_depth}) — partial result kept (not cached as complete)",
             file=sys.stderr,
         )
+        # Conservative: this marks every file in the merged chunk partial, even
+        # ones that finished cleanly during recursion. Over-marking only costs a
+        # re-extraction next run; under-marking would serve a truncated file as
+        # complete, so err toward re-extraction.
+        _mark_partial(result)
         return result
 
     print(
@@ -1939,6 +1991,10 @@ def extract_corpus_parallel(
                 # that changes _EXTRACTION_SYSTEM re-extracts instead of replaying
                 # this vintage forever (#1939).
                 prompt=_extraction_system(deep=deep_mode),
+                # A truncated/partial chunk must not be checkpointed as
+                # authoritative: pass the partial file set so its entry is
+                # stamped ``partial: True`` and re-dispatched next run.
+                partial_source_files=_partial_source_files(result) or None,
             )
         except Exception as _exc:  # noqa: BLE001 — checkpoint is best-effort
             print(f"[graphify] incremental cache checkpoint failed: {_exc}", file=sys.stderr)
diff --git a/tests/test_chunking.py b/tests/test_chunking.py
index 19e65efbc..dea9f68f2 100644
--- a/tests/test_chunking.py
+++ b/tests/test_chunking.py
@@ -329,6 +329,33 @@ def test_checkpoint_scopes_cache_writes_to_chunk_files(tmp_path):
     assert a_cache and any(n["id"] == "a_ok" for n in a_cache["nodes"])
 
 
+def test_truncated_chunk_is_cached_partial_and_missed_on_reload(tmp_path):
+    """A single-file chunk that stays truncated is checkpointed as a PARTIAL
+    entry, so reloading it is a cache miss (the file re-dispatches next run)
+    instead of serving the incomplete node set forever."""
+    from graphify.llm import extract_corpus_parallel, _extraction_system
+    from graphify.cache import load_cached
+
+    doc = tmp_path / "doc.md"; doc.write_text("# Heading\nlots of prose\n")
+
+    def truncated(chunk, **kwargs):
+        return {
+            "nodes": [{"id": "n1", "source_file": "doc.md", "file_type": "document"}],
+            "edges": [], "hyperedges": [],
+            "input_tokens": 1, "output_tokens": 1,
+            "finish_reason": "length",
+        }
+
+    with patch("graphify.llm.extract_files_direct", side_effect=truncated):
+        extract_corpus_parallel(
+            [doc], backend="kimi", root=tmp_path,
+            token_budget=None, chunk_size=1, max_concurrency=1,
+        )
+
+    # The entry was written but stamped partial, so load_cached rejects it.
+    assert load_cached(doc, tmp_path, kind="semantic", prompt=_extraction_system()) is None
+
+
 def test_checkpoint_writes_deep_namespace_in_deep_mode(tmp_path):
     """#1894: the per-chunk checkpoint must follow the run's mode — a
     deep_mode=True run checkpoints into cache/semantic-deep/, leaving the
@@ -712,6 +739,68 @@ def test_adaptive_retry_single_file_truncation_does_not_recurse(tmp_path, capsys
     assert "single-file chunk" in err and "truncated" in err
 
 
+def test_adaptive_retry_marks_single_file_truncation_partial(tmp_path):
+    """A non-splittable single-file truncation keeps its partial result but
+    marks every item ``_partial`` so it is not cached as complete."""
+    from graphify.llm import _extract_with_adaptive_retry
+
+    f = tmp_path / "huge.py"; f.write_text("x")
+
+    def stub(chunk, **kwargs):
+        return _stub_with_finish(len(chunk), finish_reason="length")
+
+    with patch("graphify.llm.extract_files_direct", side_effect=stub):
+        result = _extract_with_adaptive_retry(
+            [f], backend="kimi", api_key=None, model=None, root=tmp_path, max_depth=3
+        )
+
+    assert result["nodes"], "the partial result should still be returned"
+    assert all(n.get("_partial") for n in result["nodes"])
+
+
+def test_adaptive_retry_marks_max_depth_giveup_partial(tmp_path):
+    """When recursion caps at max_depth with everything still truncated, the
+    merged partial result is marked ``_partial`` on every item."""
+    from graphify.llm import _extract_with_adaptive_retry
+
+    files = [tmp_path / f"f{i}.py" for i in range(8)]
+    for f in files:
+        f.write_text("x")
+
+    def stub(chunk, **kwargs):
+        return _stub_with_finish(len(chunk), finish_reason="length")
+
+    with patch("graphify.llm.extract_files_direct", side_effect=stub):
+        result = _extract_with_adaptive_retry(
+            files, backend="kimi", api_key=None, model=None, root=tmp_path, max_depth=2
+        )
+
+    assert result["nodes"]
+    assert all(n.get("_partial") for n in result["nodes"])
+
+
+def test_adaptive_retry_successful_split_is_not_marked_partial(tmp_path):
+    """A truncation that IS recovered by splitting yields a complete result —
+    it must NOT carry the partial marker."""
+    from graphify.llm import _extract_with_adaptive_retry
+
+    files = [tmp_path / f"f{i}.py" for i in range(4)]
+    for f in files:
+        f.write_text("x")
+
+    def stub(chunk, **kwargs):
+        finish = "length" if len(chunk) == 4 else "stop"
+        return _stub_with_finish(len(chunk), finish_reason=finish)
+
+    with patch("graphify.llm.extract_files_direct", side_effect=stub):
+        result = _extract_with_adaptive_retry(
+            files, backend="kimi", api_key=None, model=None, root=tmp_path, max_depth=3
+        )
+
+    assert result["nodes"]
+    assert not any(n.get("_partial") for n in result["nodes"])
+
+
 def test_corpus_parallel_uses_adaptive_retry(tmp_path):
     """End-to-end: extract_corpus_parallel routes through adaptive retry,
     so a chunk that truncates gets split and merged transparently before
diff --git a/tests/test_partial_cache.py b/tests/test_partial_cache.py
new file mode 100644
index 000000000..9dd5be91d
--- /dev/null
+++ b/tests/test_partial_cache.py
@@ -0,0 +1,141 @@
+"""Tests for partial-extraction cache promotion.
+
+A truncated LLM chunk (`finish_reason="length"` that could not be recovered by
+splitting, or a max-depth adaptive-retry give-up) yields an incomplete node set.
+It is tagged with an internal ``_partial`` marker; ``save_semantic_cache`` stamps
+that file's entry ``partial: True``, and ``load_cached`` then treats a partial
+entry as a cache MISS so the file is re-dispatched instead of served forever.
+"""
+
+from graphify import llm
+from graphify.cache import (
+    save_semantic_cache,
+    load_cached,
+    _group_has_partial_marker,
+)
+
+
+def _doc(tmp_path):
+    doc = tmp_path / "doc.md"
+    doc.write_text("# Heading\nsome prose\n", encoding="utf-8")
+    return doc
+
+
+def test_intrinsic_partial_marker_makes_entry_a_cache_miss(tmp_path):
+    doc = _doc(tmp_path)
+    nodes = [{"id": "n1", "label": "Heading", "source_file": "doc.md", "_partial": True}]
+    saved = save_semantic_cache(nodes, [], root=tmp_path, prompt="P")
+    assert saved == 1
+    # The stamped entry is present on disk, but load_cached rejects it.
+    assert load_cached(doc, root=tmp_path, kind="semantic", prompt="P") is None
+
+
+def test_partial_source_files_arg_stamps_entry(tmp_path):
+    doc = _doc(tmp_path)
+    # No intrinsic marker; partial-ness comes only from the explicit arg.
+    nodes = [{"id": "n1", "label": "Heading", "source_file": "doc.md"}]
+    save_semantic_cache(nodes, [], root=tmp_path, prompt="P", partial_source_files=["doc.md"])
+    assert load_cached(doc, root=tmp_path, kind="semantic", prompt="P") is None
+
+
+def test_non_partial_entry_loads_normally(tmp_path):
+    doc = _doc(tmp_path)
+    nodes = [{"id": "n1", "label": "Heading", "source_file": "doc.md"}]
+    save_semantic_cache(nodes, [], root=tmp_path, prompt="P")
+    loaded = load_cached(doc, root=tmp_path, kind="semantic", prompt="P")
+    assert loaded is not None
+    assert len(loaded["nodes"]) == 1
+
+
+def test_partial_entry_self_heals_on_complete_reextraction(tmp_path):
+    doc = _doc(tmp_path)
+    partial = [{"id": "n1", "source_file": "doc.md", "_partial": True}]
+    save_semantic_cache(partial, [], root=tmp_path, prompt="P")
+    assert load_cached(doc, root=tmp_path, kind="semantic", prompt="P") is None
+    # A later complete extraction overwrites the same content-hash key with a
+    # non-partial entry, which then serves normally.
+    complete = [
+        {"id": "n1", "source_file": "doc.md"},
+        {"id": "n2", "source_file": "doc.md"},
+    ]
+    save_semantic_cache(complete, [], root=tmp_path, prompt="P")
+    loaded = load_cached(doc, root=tmp_path, kind="semantic", prompt="P")
+    assert loaded is not None
+    assert len(loaded["nodes"]) == 2
+
+
+def test_merge_existing_accumulates_slices_and_stays_partial(tmp_path):
+    """A file sliced across chunks: an earlier truncated slice must not be
+    dropped (nor the entry promoted to complete) by a later clean slice's
+    merge_existing checkpoint. The union keeps both slices and the entry stays
+    partial until a fully-clean re-extraction overwrites it."""
+    doc = _doc(tmp_path)
+    partial = [{"id": "n1", "source_file": "doc.md", "_partial": True}]
+    save_semantic_cache(partial, [], root=tmp_path, prompt="P")
+    fresh = [{"id": "n2", "source_file": "doc.md"}]
+    save_semantic_cache(fresh, [], root=tmp_path, prompt="P", merge_existing=True)
+    # Normal read is still a miss: the file had a truncated slice, so it must be
+    # re-dispatched rather than served.
+    assert load_cached(doc, root=tmp_path, kind="semantic", prompt="P") is None
+    # But nothing was lost — both slices are present in the accumulated entry.
+    peek = load_cached(doc, root=tmp_path, kind="semantic", prompt="P", allow_partial=True)
+    assert peek is not None
+    assert {n["id"] for n in peek["nodes"]} == {"n1", "n2"}
+
+
+def test_stamped_manifest_excludes_partial_files():
+    """A truncated file produced output this run but is left unstamped in the
+    manifest (like a failed chunk) so detect_incremental re-queues it."""
+    from pathlib import Path
+    from graphify.cli import _stamped_manifest_files
+
+    files_by_type = {"document": ["a.md", "b.md"], "code": ["x.py"]}
+    sem_result = {
+        "nodes": [
+            {"id": "1", "source_file": "a.md"},
+            {"id": "2", "source_file": "b.md"},
+        ],
+        "edges": [], "hyperedges": [],
+    }
+    out = _stamped_manifest_files(files_by_type, sem_result, Path("."),
+                                  partial_source_files={"b.md"})
+    # a.md extracted cleanly -> stamped; b.md truncated -> excluded; code kept.
+    assert out["document"] == ["a.md"]
+    assert out["code"] == ["x.py"]
+
+
+def test_group_has_partial_marker():
+    assert _group_has_partial_marker({"nodes": [{"_partial": True}]}) is True
+    assert _group_has_partial_marker({"edges": [{"_partial": True}]}) is True
+    assert _group_has_partial_marker({"nodes": [{"id": "a"}], "edges": [], "hyperedges": []}) is False
+    assert _group_has_partial_marker({}) is False
+
+
+def test_mark_partial_and_partial_source_files():
+    result = {
+        "nodes": [{"id": "a", "source_file": "x.md"}],
+        "edges": [{"source": "a", "target": "b", "source_file": "x.md"}],
+        "hyperedges": [{"id": "h", "source_file": "y.md"}],
+    }
+    llm._mark_partial(result)
+    assert result["nodes"][0]["_partial"] is True
+    assert result["edges"][0]["_partial"] is True
+    assert result["hyperedges"][0]["_partial"] is True
+    assert llm._partial_source_files(result) == ["x.md", "y.md"]
+
+
+def test_partial_source_files_empty_when_unmarked():
+    result = {"nodes": [{"id": "a", "source_file": "x.md"}], "edges": [], "hyperedges": []}
+    assert llm._partial_source_files(result) == []
+
+
+def test_strip_partial_markers_removes_internal_key():
+    result = {
+        "nodes": [{"id": "a", "_partial": True}],
+        "edges": [{"source": "a", "target": "b", "_partial": True}],
+        "hyperedges": [{"id": "h", "_partial": True}],
+    }
+    llm._strip_partial_markers(result)
+    assert "_partial" not in result["nodes"][0]
+    assert "_partial" not in result["edges"][0]
+    assert "_partial" not in result["hyperedges"][0]