fix(dedup): let the defining file win an ID collision, and warn only about real loss

[bchan84x]

Fixes the ID-collision handling described in #1851.

The problem

A node ID is <slug(source_path)>_<slug(entity)>, so a doc that merely references an entity mints that entity's own ID and collides with the entity's defining node by construction. The pre-dedup pass in deduplicate_entities handled that badly:

• Survivor was insertion-order. seen_ids kept whichever node arrived first, so chunk order decided whether an entity kept its own attributes or a passing cross-reference's thinner label.
• The warning asserted a wrong cause. It hardcoded "two files with the same name in different directories" — in the reported case the two files had different names, and nothing was lost (edges are keyed by ID and rewired to the survivor).
• The genuinely lossy drop was silent. The warning was gated on existing_sf != new_sf, so two same-ID nodes from the same file — the extractor emitting two different labels — were dropped with no output at all.

The change

graphify/dedup.py only.

1. Definer wins. When several nodes share an ID, the survivor is the one whose source_file is the file the ID encodes; otherwise first-seen, as before. _id_prefixes() accepts any trailing slice of the path, so absolute paths, repo-relative paths, and pre-Silent data loss: doc files with the same name in different directories produce colliding node IDs across extraction chunks #1504 bare-filename-stem IDs all resolve.
2. Report in proportion to what the drop costs.
   • A cross-reference folding into the node it references → silent (nothing is lost).
   • Same file, two labels for one ID → [graphify] note: naming the discarded label (this was the previously-silent case).
   • Two different files both encoding the ID → still WARNING, now stating the ID-scheme cause and keeping the extract per-subfolder + merge-graphs advice. This is the Silent data loss: doc files with the same name in different directories produce colliding node IDs across extraction chunks #1504 case, and its existing test still passes unchanged.

Tests

New tests in tests/test_dedup.py: definer-wins in both insertion orders, reference collisions stay silent and keep their edges, absolute source_file paths still resolve, same-file relabels emit the note, plus a unit test for _defines_id. Both existing collision tests (test_cross_chunk_id_collision_emits_warning, test_same_id_same_source_file_no_warning) are untouched and green.

pytest tests/test_dedup.py → 40 passed. ruff check clean.

Notes

• CHANGELOG.md deliberately not touched — the entries there look maintainer-written at release time. Happy to add one if you'd prefer.
• Two local failure clusters are environmental, not from this change: test_falkordb_integration.py (no FalkorDB module in my local Redis) and test_skillgen.py (git show of baseline blobs, which a --depth 1 clone lacks; CI uses fetch-depth: 0).

Refs #1851

[rickgonzalez]

A scan of this PR to explain it - using a tool I'm working on. Hope that it helps: https://www.lenzon.ai/viewer/cmrju75nq0011odzhizb1hp1p?voice=google-chirp3