Score ranked grounding alternatives in MedMentions

[maziyarpanahi]

Pull Request

Description

Include ranked alternatives when calculating top-k accuracy for non-abstaining GroundedSpan outputs. Withheld selections and empty selections remain abstentions; sequence providers keep their existing behavior.

This follows the report and immutable reproduction shared by Belal Embaby. The fix was reproduced against current upstream and implemented in existing tracked source/test files; the contributor's fork was not modified.

Type of Change

• Bug fix
• Test addition/improvement

Changes Made

• Score selected candidates followed by ranked alternatives, preserving UMLS filtering and the requested depth.
• Extend existing synthetic benchmark tests for rank cutoffs, explicit abstention, no selection, both provider shapes, unchanged output ordering and value-free aggregate reports.
• Update the evaluator docstring and changelog; no corpus or restricted vocabulary is bundled.

Testing

The published candidate is 593d055f63000aae6b1be47d00c11093126ed097. It retains the issue contribution and normally integrates frozen SDK prerequisite 5e635dd707ebd74c2d0806b0cb6415b53ab4d7b0 from #3971. The issue-specific scope is unchanged; shared SDK/CI/build-budget changes belong to that prerequisite.

• make format, make lint, make format-check, make type-check and git diff --check: passed, without unrelated formatting changes.
• The repository's configured type check covers four modules, not the entire codebase.
• Focused regression/public-API tests: 55 passed.
• Shared wheel/import budget and CI-wiring tests: 41 passed.
• Repository-policy and declared development-profile dependency audit: passed. This audit does not clear separately reviewed optional dependency profiles.
• High-severity/medium-confidence Bandit gate: passed, zero high-severity findings, no files skipped. Lower-severity results are not represented as zero.

Post-integration full offline pytest is pending; no full-suite pass is yet claimed for this exact head. Its local runtime is Python 3.13.11, with a private environment and unique OS temporary fixture directory. The previous complete local full receipt was for be790dc2982bb10e2b1aa871b3c8febe8ff5b077: 23,123 passed, 222 skipped, 24 warnings, zero failures/errors. That older result is historical evidence only and does not validate the new integration head.

SDK #3971 must still land and this exact pushed head must pass its own full local and hosted checks before merge. Hosted checks remain pending; this is not merge or release clearance. No tests or assertions were removed or weakened, and no skipped check or dependency waiver is introduced.

Documentation

• Updated the relevant public docstring and CHANGELOG.md.

Code Quality

• Canonical formatting/lint/type gates passed.
• Performed a separate diff review and checked accepted-input/negative controls.
• No unrelated formatting changes or new tracked files.

Dependencies

• No issue-specific runtime dependencies; shared SDK prerequisite changes are tracked in Validate installed SDK compatibility across desktop Python platforms #3971.

Checklist

• Read the contributing guide, Code of Conduct and maintainer guide.
• One focused issue implementation commit is retained; shared prerequisite integration is recorded separately. The contributor's original fork remains unchanged.

Related Issues

Closes #3569

Screenshots/Examples

Not applicable; backend/evaluation contract fix with synthetic unit coverage.

[maziyarpanahi]

Thank you for the focused contribution for #3569. Its implementation and regression coverage already landed on master through #3650, and the issue is closed. I verified the release merge is present and the corresponding regressions still pass on current master.

This PR is held as superseded; the resolved issue provides no remaining implementation scope for another merge. I am retaining the original branch and discussion. Any additional behavior should be scoped to a new maintainer-backed issue before further changes or merge.