Fuzzy dedup falsely merges similar-but-distinct entities (chip models, codenames with shared prefixes)

[myf-cpu]

Fuzzy dedup uses Jaro-Winkler string similarity (_MERGE_THRESHOLD = 92.0) and
incorrectly merges entities that are similar but semantically distinct.

Example cases:

Label A Label B Jaro-Winkler Result
───────── ───────── ───────────── ─────────────────
ASR1603 ASR1605 ~95.2 falsely merged
ASR1603 ASR1606 ~95.2 falsely merged
cranel cranelr ~96.3 falsely merged
cranel cranelrh ~95.2 falsely merged
cranelr cranelrh ~96.3 falsely merged

These are different chip name / hardware variants, not typos or formatting
variants. Merging them causes:

Nodes disappear from the graph (only the winner survives)
All edges are rewired to the survivor, creating false relationships
The merged entity appears connected to everything all variants were connected to
Do you have any suggestions regarding this issue？

[myf-cpu]

Current workaround

--dedup-skip-fuzzy KEYWORD exists and works, but has limitations:

graphify extract . --dedup-skip-fuzzy 1601 --dedup-skip-fuzzy cranel
Upon detecting these keywords, it will skip dedup
Now this method is working, but I don't know if it will bring other problems, or if you have any other better methods

[safishamsi]

Fixed in commit 6f8e6c5 (v0.8.5). Two new guards in dedup.py: _is_variant_pair() catches same-stem/different-suffix pairs like ASR1603/ASR1605, and _short_label_blocked() prevents Jaro-Winkler's prefix bonus from merging short labels unless they're same-length single-char typos. Both guards are gated at max(len) >= 12 so long labels still go through normal fuzzy matching. Five tests added in tests/test_dedup.py.