The !op.perform_delim case is always used when the system detects correlation with hard to duplicate eliminate types string, list, etc. In that case, the window function is pushed instead of doing the deduplication. In the existing test cases, it was always the case, that the window function and corresponding correlated column was added as the first element. Therefore the code just assumed, that we always have to just consider the first column.

I think your fix is sane, because we can assume that the order of correlated columns remains unchanged. As there might be multiple "delim_index"es, the linear scan should preserve the intended semantics—I think.

Thanks for the feedback! I understand the problem a bit better now, and I've created a wrapper instead:

struct CorrelatedColumns {
private:
	using container_type = vector<CorrelatedColumnInfo>;

public:
	CorrelatedColumns() : delim_index(1ULL << 63) {
	}

	void AddColumn(container_type::value_type info) {
		// Add to beginning
		correlated_columns.insert(correlated_columns.begin(), std::move(info));
		delim_index++;
	}

	void SetDelimIndexToZero() {
		delim_index = 0;
	}

	idx_t GetDelimIndex() const {
		return delim_index;
	}

	// ...

private:
	container_type correlated_columns;
	idx_t delim_index;
};

This keeps track of how the delim_index is pushed around the vector, which seems to work as it passes the test.

Thanks!