So there is no rational controversy over whether to use time-series validation or a walk-forward analysis (WFA). Some have noted that an expanding window is better than a rolling window and I find that to be true. But a walk-forward backtest can use either (expanding window or rolling window).
But here is the thing. A walk-forward backtest or time-series validation can test either a parameters OR or a single algorithm.
Here is my algorithm in this artifact with no internal parameters: For each training period find the best SMA within training data only and then test the next period (with no data leakage). Expand the window and repeat. But the algorithm has no variables or parameters.
This makes it like nested cross-validation. For nested cross-validation with more than one parameter, one would use k-fold validation in the training period allowing you to use validation to select a parameter (when you have more than one) and then use a true test sample and walk that forward.
Scikit-learn addresses your very legitimate concerns about regular validation methods here I believe: Nested versus non-nested cross-validation
Nested cross-validation has been discussed before by @jlittleton: Walk forward time series testing with inner kfold loop for AI training. I guess I was the only one to give it a heart at the time (Nov 24).