fix: handle multiple GLiNER labels share the same span

[asteier2026]

Summary

• resolve_overlaps previously used alphabetical label order as the final tiebreaker when two GLiNER detections shared the exact same character span, causing higher-confidence labels to be silently dropped in favour of lower-confidence ones.
• Added -item.score to the sort key (before item.label) so the highest-scoring label wins on exact-span ties.

Example: "Mum" was tagged as both relationship (score 0.941) and last_name (score 0.719). Before this fix, last_name won because l < r alphabetically. After, relationship correctly wins.

Test plan

• Existing test_detection_postprocess.py parametrized test covers the new behaviour — verify it passes with make test
• Spot-check a rewrite run on a text with relationship/family terms to confirm relationship entities surface correctly

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR changes exact-span entity conflict resolution and preserves detector provenance during occurrence propagation. It also extends W&B benchmark-import identity metadata and guarded SDK configuration behavior.

• Prefer higher-confidence detections when labels share an exact character span.
• Avoid creating propagation copies at coordinates already occupied by original entities.
• Add benchmark identity validation, CLI/config projection, retry refresh behavior, documentation, and tests.

Confidence Score: 4/5

The PR is not yet safe to merge because mixed-source exact-span collisions can still replace detector entities with synthetic augmented entities.

The propagation-specific provenance regression is addressed, but resolve_overlaps still compares real detector confidence against synthetic score-1.0 values from apply_augmented_entities, leaving the previously reported mixed-source entity replacement behavior reachable.

Files Needing Attention: src/anonymizer/engine/detection/postprocess.py

Important Files Changed

Filename	Overview
src/anonymizer/engine/detection/postprocess.py	Adds score-based exact-span ordering and skips propagation copies at original entity coordinates.
tests/engine/test_detection_postprocess.py	Adds coverage for exact-span score selection and preservation of detector provenance during propagation.
tools/measurement/import_wandb_run.py	Adds benchmark identity inputs to strict W&B imports.
tools/measurement/measurement_tools/wandb_models.py	Extends typed W&B configuration models with validated benchmark identity metadata.
tools/measurement/measurement_tools/wandb_setup.py	Updates guarded W&B environment and publication configuration handling.

_{Reviews (4): Last reviewed commit: "fix: skip propagation copies for spans a..." | Re-trigger Greptile}

[asteier2026]

Changes made per Greptile suggestions