Follow-up with a small test, so this isn't just opinion. I wrote 4 sample clue types in the style your book might use and checked each against what I can do, the way a reader pasting text into a chatbot would.

1. Acrostic (first letters of the 12 paragraphs of a chapter spell a word). Solved instantly. Any hidden-letter pattern in plain text is the first thing a model checks.
2. Historical trivia ("the number of BTC in the pizza order, times the block height of the first halving"). Solved instantly, 10,000 × 210,000. Bitcoin history is exactly what models know best, and that's your whole book.
3. Physical artifact ("the third word on the page whose number is the year Satoshi went silent, in the first printed edition"). Not solvable from a pasted text: page numbers depend on the print layout. It costs a solver a copy of the book, which also answers your question 3.
4. Cross-chapter with no shared marker (a fragment in ch. 4 that only makes sense combined with a detail in ch. 19, with nothing linking them). This is the hard one for models too: with two long chapters and nothing marking the link, I'd have to test every pair of details. I can't prove that on a toy example, so treat it as the type to red-team with your real text. It's the one that scales to months.

My practical suggestion: build the key from type 3 and 4 clues, and use 1 and 2 only as on-ramps, so people get early wins and keep going. Before you publish, run every clue through 2-3 current models. If any one solves it, demote it to an on-ramp.

My offer stands: send me the real clues (privately, or through a hash-committed list) and I'll do that red-team pass for free.