Converge calibration evidence-strength selection between VA-Spec and the serving layer

[bencap]

Summary

Calibration selection logic is scattered across the codebase with subtly different strategies. The serving layer's display-default selection has now been consolidated into lib/score_calibrations.py (calibration_preference_key + classification_evidence_strength), shared by lib/variant_detail.py and lib/allele_measurements.py. The VA-Spec annotation path still selects its evidence calibration/classification independently. This issue tracks converging the two so there is a single source of truth for "how strong is a classification" and "which calibration do we pick."

Why this matters

There are two legitimately different preference policies, and that's fine:

• Display default — the UI cascade (primary → investigator_provided → non-research_use_only), now calibration_preference_key. What a card/page shows by default.
• Strongest evidence — VA-Spec picks the calibration/classification with the strongest evidence for a statement (lib/annotation/statement.py, lib/annotation/util.py).

The risk is not that these differ — it's that the "evidence strength" primitive is defined twice. The serving layer now computes it in classification_evidence_strength (ACMG points → |ln(oddspaths_ratio)| → functional call, with a pathogenic/benign direction). If VA-Spec computes "strongest" by a different rule, the UI and the emitted VA statements can disagree about which evidence is strongest for the same variant — a confusing, hard-to-spot inconsistency in a clinically-sensitive surface.

Scope

• Audit VA-Spec's calibration/classification selection (lib/annotation/statement.py, lib/annotation/util.py::score_calibration_may_be_used_for_annotation and callers) against the new shared helpers.
• Have VA-Spec's "strongest evidence" selection reuse classification_evidence_strength (or, if it must differ, document why in both places and add a test pinning the intended divergence).
• Consider promoting a shared strongest_classification(...) selector and keeping the eligibility filter (score_calibration_may_be_used_for_annotation) as the separate, already-centralized concern.
• Confirm the display-default vs strongest-evidence split is intentional and documented (they should stay two named policies, not collapse).

Pointers

• Shared helpers: src/mavedb/lib/score_calibrations.py — calibration_preference_key, classification_evidence_strength
• Serving-layer consumers: src/mavedb/lib/variant_detail.py, src/mavedb/lib/allele_measurements.py
• VA-Spec selection: src/mavedb/lib/annotation/statement.py, src/mavedb/lib/annotation/util.py