# fix(dedup): merge same-ID node attributes instead of discarding them (#2091)

[SinghAman21]

What

On the default dedup=True path, when two nodes share an ID the pre-dedup step
kept one whole node dict and threw the other away. In the AST↔semantic case
(where _semantic_id_remap forces the AST node and the LLM node onto one ID so
they merge), the LLM's summary / confidence_score / references were
silently dropped — even though build()'s docstring promises a key-wise union.

The user already paid API credits for those summaries, and the only output was a
note: about the label, never mentioning the lost attributes.

How

The survivor now gap-fills from the loser instead of replacing it wholesale:

• Only keys the survivor lacks are copied over (setdefault).
• Conflicting keys (label, source_file) still resolve to the survivor, so
  _collision_rank ordering and the collision reporting are unchanged.
• The first insert now copies (dict(node)) so caller-owned node dicts are never
  mutated.

Before: dedup=True → summary=None. After: dedup=True → summary preserved,
survivor's label still wins the conflict.

Files

• graphify/dedup.py — gap-fill merge in the same-ID pre-dedup loop.
• tests/test_dedup.py — 3 tests: gap-fill preserves disjoint attrs,
  survivor wins on conflict, input nodes are not mutated.
• CHANGELOG.md — entry under 0.9.24 (unreleased).

Reference

• Issue Default dedup=True discards a same-ID node's attributes instead of merging them — LLM semantic summaries are silently lost #2091
• Root cause: graphify/dedup.py same-ID collapse loop.

Conclusion

dedup=True now yields at least the attribute set dedup=False does, with the
survivor winning genuine conflicts. No public API change; full suite green
(pre-existing unrelated test_ollama_retry_cap.py failures aside).