[E09] EF≤40% calibration artifacts

[duckyquang]

Task metadata

Field	Value
Task ID	E09
Category	Evaluation
Priority	Medium
Complexity	Low
Est. days	1
Blocked by	M10, E07

Features / method

Produce the EF<=40% calibration artifacts that E04 never emitted.

Description

Same situation as E08: E04 (#10) is closed, src/primed_ai/evaluation/calibration.py implements Platt scaling from continuous LVEF to EF<=40% probabilities, and there is no calibration output in the repo.

A deployment-risk paper that reports AUROC without calibration is only telling half the story. AUROC is rank-based and says nothing about whether the probabilities are usable as a risk score at the bedside.

Steps:

1. Fit the Platt scaler on val predictions, apply to test. Do not fit on test.
2. Report expected calibration error and a reliability diagram for the full condition.
3. Repeat under echo_dropped and ecg_dropped — the question is whether the model stays honest about its own confidence when a modality goes missing, or gets overconfident.
4. Write the results into results/ with provenance.

Acceptance criteria:

• ECE and reliability diagram for all three conditions.
• Scaler fit on val only, stated explicitly in the output.
• Result file records checkpoint, seed, and prediction source.