inclusion-numerique / coop-mediation-numerique / 58e50680-d186-484b-8bc8-04b1f5fb0952

Pull #497

circleci

feat: improve structure fusion scoring and review export

Significantly reduces the manual review burden by detecting more
true duplicates automatically and avoiding false positives.

Scoring improvements (detect-duplicate-structures, generate-structures-action-plan):
- Treat clusters of type 'mixte' like 'doublon_certain' with per-pair
  scoring (instead of bulk verification_manuelle), uncovering hundreds
  of auto/probable fusions previously hidden in mixed clusters.
- Boost address score to 1.0 when one normalized address is contained
  in the other (e.g. "Lupino" vs "LUPINO PARVIS NOTRE DAME VICTOIRE").
- Add address abbreviations: VC (voie communale), RT (route), ZA, ZI, CH.
- Redistribute geo weight when coords are unavailable, OR when address
  strongly indicates the same place (>=0.85): prevents penalizing
  structures with missing or erroneous coords.
- Normalize "commune de/du", "mairie de/du", "ville de/du" to a single
  "ville" canonical token so variants match.
- Detect "service keywords" (EPN, médiathèque, CCAS, France services,
  MJC, etc.): when one name has such a keyword and the other does not,
  they are distinct entities even with shared SIRET/address. Disables
  the address-contained heuristic and keeps geo in the score.

Sync resilience (findOrCreateStructure):
- After strict siret+codeInsee miss, fall back to siret-only with
  normalized contained-name match. This catches Dataspace structures
  whose codeInsee diverges from the coop's, without merging an EPN
  with its parent town hall (asymmetric-service-keyword guard).

Review output:
- generate-structures-action-plan: structures-fusion-review.csv now
  uses cluster-grouped format (CIBLE + sources + empty line between
  clusters, sorted by ascending score), matching the existing format
  Tim uses for his manual reviews.
- export-duplicate-sirets: cluster-grouped CSV (empty line between
  SIRETs) and exclude empty-string siret. Enrich each row with
  nom_api, adresse_api, corre... (continued)

688 of 10878 branches covered (6.32%)

Branch coverage included in aggregate %.

26 of 150 new or added lines in 4 files covered. (17.33%)

2111 of 17142 relevant lines covered (12.31%)

1.94 hits per line