save_manifest drops untouched files on incremental runs, forcing full re-extraction

[dkauto2026-cyber]

Summary

save_manifest() in graphify/detect.py rebuilds the manifest only from the files passed to it. On an incremental extract/update run that processes a small subset of the corpus, every other file's entry is dropped from the manifest.

The next detect_incremental() then sees those files as missing from the manifest and flags the entire corpus as new — forcing a full, expensive re-extraction.

Where

graphify/detect.py, save_manifest() (currently ~line 843 on v8):

existing = load_manifest(manifest_path)
manifest: dict[str, dict] = {}          # <-- starts empty
for file_list in files.values():
    for f in file_list:
        ...
        manifest[f] = entry             # <-- only this run's files land here

existing is loaded but used solely for per-file prev lookups; it never seeds the output manifest.

Reproduce

1. Run a full build so the manifest has the whole corpus.
2. Run an incremental update touching only a few files.
3. Run detect_incremental() again — it reports the whole corpus as new.

Fix

Seed manifest from the existing on-disk manifest (pruned to files that still exist) before the loop. Deleted files are still dropped, so deletions self-heal. PR attached.

[safishamsi]

Fixed in v0.8.10 — save_manifest now seeds the output manifest from the existing on-disk manifest before the loop, pruned to files that still exist on disk. Incremental callers passing a subset of files no longer erase entries for untouched files.