Checked against the current Windows build. Where an article names a menu, a button or a message, that is the wording in the application, not a paraphrase.
Which of the two you have, in one command
creasepoint objects scan.pdf --page 1 prints the page's size in points and then every object on it. When the page carries a rotation attribute, the first line says so — shown turned 90° — followed by a note that the bounds listed underneath are in the page's own frame and not the turned one. No such line means the page carries no rotation, and a page with no rotation that still looks sideways is sideways because what is drawn on it is.
The distinction matters less than people expect and more than not at all. A scanned landscape page is very often a portrait sheet with a rotation of 90 degrees on it, and that is normal rather than a defect. What differs between the two cases is the shape of the page you finish with, which the next section gets to.
Rotate changes an attribute; nothing on the page moves
Rotate sits on the Pages tab and on Home. Click a page's number to select it and shift-click another for a range, or select nothing and the whole document turns. The panel offers 90° right, 180° and 90° left, and each turn is relative to the page's current rotation, so pressing it twice is one turn of 180 rather than an argument about absolute angles. From a script: creasepoint rotate scan.pdf --pages 1,3 --turns 1 --out out.pdf.
What it writes is the page's rotation entry. Nothing in the content is moved and no image is re-encoded, which is why turning a scan costs it no quality whatever. Rotation does rebuild the document, though, so read the notes underneath the result: attachments, metadata and bookmarks are deliberately carried back across, and anything that could not be is named there rather than lost in silence.
OCR reads what the page draws, so rotate first
Making a scan searchable renders each page at 300 dpi, recognises the printed words in that rendering, and writes them back as one invisible text object per word, stretched to the box that word occupies. The rendering is the page as a reader shows it, upright frame, after the rotation attribute has been applied. A page that still looks sideways to you looks sideways to the recogniser.
So the sequence is fixed. Rotate, look at the pages, then run OCR. The other way round gives you a text layer built from a bad reading, and rotating afterwards turns that layer without re-running recognition. Nothing in the result will warn you either: the counts report pages recognised, pages skipped and words placed, and a page of confident nonsense counts as words placed. Search the output for a phrase you can see on the page before you rely on any of it.
The order for everything else
Rotate, then OCR, then whatever changes which pages are in the file — extracting, deleting, reordering, merging — then page numbers and headers, and compression last of all. Numbering before the pages are settled produces a page 7 that follows page 4. Compressing before the rest means doing it twice, because every later step changes the size.
One thing works in your favour here. Numbers, headers and watermarks are laid out in the page's upright frame, the box a reader actually shows, so the bottom right is the bottom right of what you see even on a page carrying a rotation attribute. Two things do not. OCR refuses a password-protected document until the password has been removed. And Skip pages that already have text should stay ticked for a document that is part typed and part scanned; on the command line that is the default, and --force is what turns it off.
What rotation will not fix
A page fed through the scanner four degrees off square is not a rotation problem. Rotate does quarter turns and nothing else, there is no deskew here, and OCR does not straighten a page before reading it. A crooked scan reads worse than a straight one, and the only reliable fix is to scan it again.
The same goes for a bundle where half the pages face one way and half the other. Select the pages that need turning by their numbers and turn only those; the selection exists so that a mixed document does not have to be split first. Then check the result at 100 per cent rather than at whatever zoom made everything fit on screen — Ctrl+0 puts it back.
The two places worth looking next.
Read the matching guide if you want the reasoning rather than the fix, and check limitations before spending any longer on it — some of what people look for here is absent on purpose.