v0.9.27 incremental --no-cluster preserves nodes but loses and mis-targets cross-file edges after #2169 #2213

Summary

The v0.9.27 fix for #2169 prevents the catastrophic changed-file-subset overwrite, but incremental code-only, no-cluster extraction still does not produce the same graph as a forced full extraction of the identical source tree.

In a minimal public corpus, an edit that only appends a Python comment preserves all 218 nodes but leaves the incremental graph with 575 edges versus 577 from a forced rebuild. Three canonical cross-file edges are missing and one absolute-path-shaped import target appears instead.

Environment

• Graphify: 0.9.27
• Python: 3.11.15
• NetworkX: 3.6.1
• OS: Ubuntu 26.04 on WSL2 (Linux 6.18.33.2-microsoft-standard-WSL2, x86_64)
• Target corpus: public pallets/itsdangerous, commit 672971d66a2ef9f85151e53283113f33d642dabd
• Extraction mode: --code-only --no-cluster, with a separate --out root

Reproduction

root="$(mktemp -d)"
git clone https://github.com/pallets/itsdangerous.git "$root/target"

# Baseline.
HOME="$root/home" XDG_CACHE_HOME="$root/home/.cache" \
  uvx --from graphifyy==0.9.27 graphify extract "$root/target" \
  --code-only --no-cluster --out "$root/out" --force

# Change one file without changing its AST.
printf '\n# Graphify v0.9.27 incremental-rebuild verification marker.\n' \
  >> "$root/target/src/itsdangerous/encoding.py"

# Incremental extraction.
HOME="$root/home" XDG_CACHE_HOME="$root/home/.cache" \
  uvx --from graphifyy==0.9.27 graphify extract "$root/target" \
  --code-only --no-cluster --out "$root/out"

# Forced full extraction of the exact same changed source tree.
HOME="$root/home" XDG_CACHE_HOME="$root/home/.cache" \
  uvx --from graphifyy==0.9.27 graphify extract "$root/target" \
  --code-only --no-cluster --out "$root/full" --force

Compare node and edge dictionaries as order-independent JSON sets.

Observed result

Incremental scan: 1 changed, 17 unchanged, 0 deleted
Incremental graph: 218 nodes, 575 edges
Forced full graph: 218 nodes, 577 edges
Node sets equal: True
Edge sets equal: False

Incremental graph SHA-256:

0edc5402423eee96db8c157ef359043721f85fd379679a37401dcde02d4af29d

Forced-full graph SHA-256:

e2436f2d8342e12f4d4ba9a58c0d374ff7a419705a5fb364f7ec0c58406809e4

Missing from the incremental graph:

src_itsdangerous_encoding_base64_decode
  --calls--> src_itsdangerous_exc_baddata

src_itsdangerous_encoding
  --imports--> src_itsdangerous_exc_baddata

src_itsdangerous_encoding
  --imports_from--> src_itsdangerous_exc

The incremental graph instead adds:

src_itsdangerous_encoding
  --imports_from--> tmp_graphify_v0927_verify_target_src_itsdangerous_exc_py

The affected edges all originate from src/itsdangerous/encoding.py. The changed source only gained a comment after the final function, so these import/call relationships should be identical.

Expected result

For a change that does not alter the AST, incremental extraction and a forced full extraction of the same source tree should have equivalent node and edge sets. Cross-file targets should remain canonical rather than becoming absolute-path-shaped IDs.

Relation to #2169

#2169 was closed as fixed in v0.9.27, with the release note stating that incremental --no-cluster now:

• merges the existing graph forward; and
• canonicalises cross-file edge targets.

The first part is materially improved: unchanged nodes are no longer discarded. The reproduction above shows a residual failure in the second part and a remaining incremental/full edge-set mismatch.

Suggested regression test

1. Build a two-module corpus where module A imports and calls a symbol from module B.
2. Force-extract the full graph.
3. Append a comment to module A.
4. Incrementally extract.
5. Force-extract the same changed tree to another output root.
6. Assert equality of normalised node and edge sets, including canonical cross-file targets.