Incremental --no-cluster extraction overwrites the full graph with the changed-file subset #2169

Summary

After a successful full graphify extract, editing one code file and rerunning graphify extract . --code-only --no-cluster overwrites graphify-out/graph.json with only the changed file’s incremental extraction.

Unchanged nodes and edges disappear, and retained edges can point to absent endpoint nodes. --force restores the full graph.

This reproduces on upstream v8 (2fa6cd3) and PR #2134 commit 36ecb25. The PR author independently confirmed the same behavior on an unrelated 789-file TypeScript/Next.js corpus: #2134 (comment)

Environment

• graphifyy 0.9.25
• macOS
• TypeScript / TSX
• code-only extraction
• reproduced with and without multigraph; the minimal case below uses the default simple/raw --no-cluster path

Minimal reproduction

.
├── app/add/scan.tsx
└── src/components/ScanScreen.tsx

// app/add/scan.tsx
import { ScanScreen } from '../../src/components/ScanScreen';

export default ScanScreen;

// src/components/ScanScreen.tsx
export function ScanScreen() {
  return null;
}

graphify extract . --code-only --no-cluster
# 3 nodes, 4 edges:
# imports_from, imports, contains, re_exports

printf '\n// harmless edit\n' >> app/add/scan.tsx

graphify extract . --code-only --no-cluster
# incremental scan: 1 changed, 1 unchanged
# 1 node, 3 edges

graphify extract . --code-only --no-cluster --force
# restored: 3 nodes, 4 edges

Observed result

After the incremental run:

• only app_add_scan remains;
• the unchanged ScanScreen.tsx file/function nodes are gone;
• its contains edge is gone;
• the three retained import/re-export edges target absent, absolute-path-shaped IDs.

Representative missing targets:

private_tmp_..._src_components_scanscreen_tsx
private_tmp_..._src_components_scanscreen_scanscreen

The output is therefore both incomplete and structurally invalid.

Larger-corpus confirmation

On an Expo/React Native TypeScript corpus:

• clustered simple: 1,143 → 1,140 edges after editing app/add/scan.tsx;
• clustered edit of another TSX file: 1,143 → 1,134 edges;
• --no-cluster --multigraph: 556 nodes / 1,143 edges → 1 node / 0 edges;
• --force restores the baseline.

The PR author reproduced the same pattern on a 789-file TypeScript/Next.js corpus:

• simple: 13,001 → 12,995 edges;
• --no-cluster: 4,586 nodes / 15,340 edges → 9 nodes / 22 edges;
• multigraph: 13,240 → 13,234 edges;
• --force restores all baselines.

Expected behavior

Incremental extraction should merge or replace the changed file’s freshly extracted content in the existing graph while preserving unchanged nodes and edges.

For --no-cluster, it must not serialize the changed subset over the full graph.json.

Changed-file endpoint IDs should also be normalized against the extraction root so newly extracted import edges resolve to the unchanged target nodes.

Likely area

The --no-cluster path appears to write the current incremental merged payload directly. The carry-forward path currently appears limited to format conversion, while ordinary incremental changed-file runs do not carry the existing graph before replacement/deduplication.

Related

• extract no-op incremental run overwrites graph with empty graph #1347 addressed no-op incremental overwrites, but this is a changed-file run on current v8.
• hook-style relative changed_paths corrupt incremental code graph #1348 addressed hook-style relative paths, but this reproduces through ordinary graphify extract.
• Found while validating feat(cluster): cluster graphs — link multiple repos into one connected graph (+ multigraph mode) #2134; the multigraph feature itself is not the cause.