fix(metrics): count num_scalings without the off-by-one from shift(-1) - #139

Fixes #136.

Problem

calculate_metrics in src/vasim/simulator/analysis/plot_utils.py counted scaling events with a backward shift:

num_changes = (merged["CURR_LIMIT"] != merged["CURR_LIMIT"].shift(-1)).sum()

shift(-1) leaves the last element NaN, and NaN != x is always True in pandas, so the final row is always counted as a scaling event — num_scalings is inflated by exactly 1 on every run (a configuration that never scales reports 1).

This isn't only a displayed value: num_scalings feeds ParetoFrontier normalization (num_scalings / num_scalings.max()) and the Pareto ranking, so tuning comparisons between configurations are skewed by it.

Fix

• plot_utils.py — count against shift() and subtract the single phantom leading change (the first row has no predecessor, so NaN != value is always True). The existing empty-input guard means this never goes negative.
• ParetoFrontier.preprocess_df — num_scalings_norm was created only inside the num_scalings.max() > 0 guard but used unconditionally on the next line. Today that line is unreachable because of the off-by-one (num_scalings was never 0); fixing the count makes it reachable for any sweep where no configuration scales, which would raise KeyError. The percentile filter is moved inside the guard.

Tests

• New tests/test_plot_utils_metrics.py with fast, deterministic unit tests: the change-count truth table from the issue, the empty-input guard, and the no-scaling preprocess_df path (regression for the KeyError).
• Four e2e expectations that hardcoded the inflated counts are decremented by 1 (466→465, 347→346, 109→108, 133→132), each confirmed by running the simulator.

Validation

tests/test_plot_utils_metrics.py .....            # unit tests
tests/test_e2e_single_run_sim.py ..               # updated expectations
tests/test_e2e_multi_run_tune_with_strategy.py ...

black, isort, flake8, and pylint pass on the changed files.