Race three predictability measures; the literature's fix loses

[jadoon200]

Q13 said the shipped diagnostic does not generalise. The forecastability literature gave a
reason and a fix — a single lagged correlation reads one frequency, spectral entropy reads
all of them. The fix was tried and it lost.

Result

43 serverless + 5 fleet workloads, same target (did forecasting pay by total economic cost),
threshold-free comparison by AUC, 20,000-draw permutation test:

measure	AUC @ 6 h	p	AUC @ 12 h	p
daily autocorrelation (shipped)	0.539	0.67	0.438	0.50
spectral predictability (the field's)	0.396	0.26	0.456	0.63
low-frequency power (ours)	0.595	0.31	0.700	0.027

Spectral entropy failed — at or below a coin flip on serverless, anti-predictive on
fleet. Predicting it would rescue the diagnostic was wrong, and that is recorded.

The only measure with signal is not from the literature. It is the share of spectral
power at periods longer than the commitment window, motivated by this project's own Q1
result: a decision fixed for N hours can only exploit structure slower than N hours. Alone
among the three it is a function of the decision horizon rather than the series alone.

Why the 0.700 is not shipped

Six tests were run. Bonferroni wants p < 0.0083; the permutation p is 0.027. Picking the
largest of six AUCs and quoting its uncorrected p-value is exactly how a result gets
manufactured, and the project's own doctrine says an implausibly clean number is an artefact
until proven otherwise.

Recorded as a lead, not a finding. The follow-up is named: one pre-registered hypothesis,
fresh cohort, horizon fixed in advance. The shipped diagnostic is unchanged and keeps the
scope caveat Q13 forced on it.

Method notes

• Comparison by AUC, not each measure's own best threshold — fitting a cutoff to 43 points
  would let any measure win. Best-threshold accuracy is printed as a labelled optimistic bound.
• Numpy only, so the diagnostic keeps no transitive scipy import.
• The periodogram is Welch-averaged; a raw one is an inconsistent estimator whose entropy is
  dominated by noise.
• Tests pin the measures against signals with known answers, including two equally-predictable
  series that spectral entropy cannot separate and the window-relative measure can.

Also

Folds the autocorrelation definition into one place. It existed separately in the API and the
evaluation script — the same drift risk that let the shipped verdict contradict docs/EVAL.md
for a day.

Gate

ruff clean; mypy strict clean; 190 tests passing (was 181).