[David Bustos Usta (dfbustosus)]

Closes #1619.

Problem

_iteration_history_records in splink/internals/em_training_session.py stamped every iteration's records with self.core_model_settings.probability_two_random_records_match — the final converged value — instead of the probability stored in that iteration's entry of _core_model_settings_history:

for iteration, core_model_settings in enumerate(self._core_model_settings_history):
    records = [
        ModelParameterIterationDetailedRecord.from_settings_param_detailed_record(
            r,
            probability_two_random_records_match=self.core_model_settings.probability_two_random_records_match,  # ← always the converged value
            iteration=iteration,
        )
        ...

Effect: the match-weights and m/u interactive history chart tooltips reported the converged λ for every iteration, while probability_two_random_records_match_iteration_chart (fed by _lambda_history_records, which already reads the per-iteration value) was correct — exactly the inconsistency reported in #1619 / the linked discussion.

Fix

One line: read probability_two_random_records_match from the per-iteration core_model_settings in the loop, as suggested by Robin Linacre (@RobinL) in the discussion.

Tests

New tests/test_em_iteration_history_records.py:

• a 3-iteration history [0.001, 0.002, 0.004] must produce records carrying each iteration's own probability (fails on the old code — all records got 0.004)
• single-iteration sanity check

The session is built with EMTrainingSession.__new__ plus a stubbed settings history, since the property only reads _core_model_settings_history — no database needed.

Verification

• pytest tests/test_em_iteration_history_records.py tests/test_expectation_maximisation.py → 6 passed, 3 deselected (the real EM-session tests still pass with the change)
• ruff check clean on both files