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

The PR changes exact-span overlap resolution to prefer the highest-scoring GLiNER label and adds a regression test for that behavior. It also extends sealed W&B imports with benchmark identity metadata and updates credential handling.

• Adds score precedence for equal-length, equal-position entity spans.
• Adds typed benchmark identity fields and CLI options for W&B imports.
• Preserves launcher-provided WANDB_API_KEY inside the guarded W&B environment.
• Expands tests and documentation for the new measurement behavior.

Confidence Score: 4/5

The PR is not yet safe to merge because occurrence expansion still replaces original detector provenance with synthetic propagation metadata.

The score-based overlap ordering applies to both detector results and score-1.0 propagation copies; when expansion recreates an original occurrence, the synthetic copy wins and its source and score are emitted in final_entities.

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

Important Files Changed

Filename	Overview
src/anonymizer/engine/detection/postprocess.py	Adds confidence-based exact-span selection, but the same precedence still lets synthetic propagation copies replace detector provenance.
tests/engine/test_detection_postprocess.py	Covers the intended detector-label confidence tiebreak but does not verify preservation of detector score and source during occurrence expansion.
tools/measurement/import_wandb_run.py	Adds validated benchmark identity CLI fields and carries them into sealed-import publication payloads.
tools/measurement/measurement_tools/wandb_models.py	Defines benchmark identity validation, comparer projections, and outbound field policies.
tools/measurement/measurement_tools/wandb_setup.py	Updates guarded W&B publication behavior to retain launcher-provided API credentials.
tools/measurement/README.md	Documents benchmark identity options and environment-based W&B authentication.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  D[Detector entity<br/>real score and detector source] --> E[Expand occurrences]
  E --> P[Propagation copy<br/>score 1.0 and propagation source]
  D --> R[Resolve overlaps]
  P --> R
  R -->|score tiebreak| F[Final entity uses propagation provenance]

Loading

_{Reviews (3): Last reviewed commit: "fix: prefer highest-scoring label when m..." | Re-trigger Greptile}