Incremental merge: stray output misattributed to non-dispatched files silently deletes their prior graph contributions

[pureoption]

Summary

During incremental extraction, LLM output whose source_file is misattributed to a file that was not dispatched silently deletes that file's entire prior contribution from graph.json. The semantic cache is protected against this; the graph build is not.

Environment

graphifyy 0.9.16, Python 3.12 (WSL2 Ubuntu 24.04), gemini backend, markdown-only corpus (26 docs).

Mechanism

1. Extracting file A, the model also emits a few stray nodes/edges attributed to other files (observed: one file's extraction produced strays attributed to 7 other files, plus nonexistent forward-reference paths like src/foo.ts).
2. _save_semantic_cache(..., allowed_source_files=uncached_paths) correctly discards them from the cache (semantic cache skipped out-of-scope source_file warning) — the cache stays clean.
3. But in cli.py, the same unfiltered fresh result is extended into sem_result and reaches build_merge().
4. build_merge() derives its replace-set from the source_files present in the new chunks — so a 2-node stray fragment for file B replaces B's entire prior contribution (observed: a file whose cache entry held 20 nodes was reduced to 1 node in graph.json).
5. B's manifest entry still says unchanged, so B is never re-dispatched and its clean cache entry is never re-applied. The loss is permanent until a full rebuild.
6. Strays attributed to nonexistent files become phantom nodes carried forward forever — never in any changed set, never matched by the delete pass (observed: 24 phantoms accumulated).

Reproduction sketch

Corpus of N cross-referencing markdown docs. Force one file to re-extract (touch content). If the model attributes any output to a second, unchanged file, that file's graph contribution is replaced by the fragment. Diff unique node ids in cache/semantic/*.json against graph.json to observe the divergence.

Suggested fix

Apply the same allowed_source_files scope filter to fresh before extending sem_result (mirror of the cache-write guard). We run 0.9.16 with exactly that local patch: a run that previously clobbered 7 files holds every node, and the filter logs 9–33 dropped strays per routine run on our corpus — misattribution is common enough that the graph-side guard matters.