What word search covers on Amazon have in common, and the cover generator rebuilt around it

Done 05:20Z: no awesome-list route on GitHub. The README is the only outside page besides Reddit that has sent a person (09-26), so I looked for curated GitHub lists where a link to Puzzle Press would be live on a page GitHub already serves. There are two that fit, and neither is usable:

• ibrahimessam/Awesome-Free-Web-Tools ranks first on GitHub for "kdp interior", but it lists only its maintainer's own DoItToolz tools, a competing KDP tool hub. It's not an open list.
• brandonhimpfen/awesome-self-publishing fits on topic, and its CONTRIBUTING.md rules us out. Under "What We Don't Accept" it says "Spam, self-promotion, or low-quality content", and it also says "early-stage projects, or resources added primarily for promotion may be declined". A pull request from the maker of a three-week-old tool is that, so I won't propose one.

The other lists found were off-topic (dataviz, Kubernetes, a "KDP" that is a data platform). Don't reopen this unless someone other than us lists the tool. Last 24h: 0 real people, 0 checkouts. The repo description is unchanged; the "generator" wording is with the boss (09-27 03:05Z).

Done 05:45Z: every request for the boss is now in one file. Every remaining lever that moves visitors is on the boss's accounts, and the requests were spread across five chat replies. plan/youtube-studio-checklist.md now ends with a "Not YouTube" section, ordered by minutes to do. It has five items: the GitHub description (exact text), the GitHub traffic numbers, two spam comments on dev.to (both linked), ShelfCall yes/no, and the two Search Console query rows. The dev.to check found the second comment the log mentioned: "grok_survivor…" on the Pinterest post, alongside "devsupport" on 09-27's post. Both are still up. From the same comment thread (listwright, 09-24): Reddit removes comments whose links go through a redirect, citing "link-masking or redirecting". So if Reddit ever reopens, the link goes straight to the page, not /go/reddit.

Done 07:10Z: demo GIF re-recorded. The README and the home page both show demo.gif, and its cover frame still had the old back: the same puzzle as the front. Since 09-27 07:12 the back has a second puzzle from the book on its own card, which the README text already describes. I re-recorded it against the live site (node scripts/demo.mjs <live url>), and it now matches. The front-cover strip (cover-row.jpg) rebuilt byte-identical, so it was already current. demo.mjs's comment still said the cover "is paid"; it now says why the demo runs licensed. launch-assets and npm test (113/0) pass. It's deployed, and the live GIF's md5 matches local.

Done 07:40Z: covers rebuilt to look like the word search books that sell. The boss: "It's a word search the covers need fun". I compared ours with the first 24 Amazon results for "word search books for adults" and "large print word search". Our cover was the plain one on the page. Of the 19 comparable covers, 10 have a heavy title in two or three colours, 11 a loud letter-grid background, and 10 a number sticker. The cover now has all of those: the title in Lilita One (OFL; non-Latin titles fall back to bold), highlight loops in five colours on the front puzzle and across the word search background, a 3° tilted card and a puzzle-count starburst. coversafe (198 covers), coverink, pdfcheck, pdftext and npm test all pass. fontwarm's typed-out "2 fonts" went stale and now counts public/fonts/. Samples, cover-row and demo.gif are regenerated and deployed, with md5s matching live. The six interior samples differed only in their dates, so I kept the committed ones. Details in actual/2026-09-28.md.

Done 08:25Z: the link-preview card shows the new covers, and perf.mjs is green for the first time since 09-13. social-card.png is the image every shared link shows (dev.to, the zoo, GitHub). It still had the old plain covers, with sudoku in front. It is rebuilt with word search on top, and the ?v= on its four pages moved to 2026-09-28 so link-preview caches refetch it. The hero image was also a day stale: it still showed the answer page from before the 7pt change, and it is regenerated too. test/perf.mjs had failed the same two checks at least since 09-21. They come from the idle warm-up of pdf-lib, which shipped on 09-13. The 09-21 log noted them as "predate this" and moved on, so the test could no longer catch a real regression. It now counts the landing (118 KB, budget 200) separately from the warm-up (520 KB). It reads the warm-up's files from the built heavy-*.js, and they must start after the load event; a mutation that moves the load event makes it fail. Last 24h: no real visitor ran the app.

Done 09:05Z: two cover bugs found by rendering the cases a sample never shows. I rendered eight covers the samples never show: one puzzle, 200 puzzles with a long title, subtitle, long author and large print, every type, a Cyrillic title, and four trims. Two problems turned up. A one-puzzle book said "1 PUZZLES", on both the new sticker and the older strip. A long title ended on a stranded word ("…FOR SENIORS: VOLUME" / "3"). The count is now singular for one. The title is balanced: at the size already chosen, it wraps to the narrowest measure that keeps the same number of lines. Of the six samples, two titles now wrap differently: "LARGE PRINT / GARDEN / WORD SEARCH" (better, since "word search" stays together) and "MAZES / FOR RAINY / DAYS". The other four are identical. A new test in cover.test.js checks both fixes; with the balancing removed, it fails. npm test 114/0; coversafe, coverink, pdfcheck, browser, mobile, freecover and launch-assets all pass. The two samples, cover-row and social-card are regenerated, and their live md5s match local. The sticker overlapping the first sudoku cells is deliberate: that's how the stickers on the Amazon covers sit.

Done 09:45Z: Show HN post brought up to date, and the one question it needs is on the checklist. HN blocked the 09-16 attempt with a gate it called "temporarily restricting Show HNs", because the account was new. Nobody has checked since. The prepared post had gone stale. It said "running about a week" and "Product Hunt yesterday", it gave two different clue counts (2,128 and 2,278; the tool has 2,278), and its page-weight reply described the PDF engine as loading on click, which stopped being true on 09-13. The text is rewritten with 09-28 figures, and the old version is kept below it marked superseded. The failures it cites are ones the log already records. The numbers that move are marked † to re-check on the day. The boss's checklist gains one item: try the real post once. If the gate is still up, nothing is posted. Don't submit a test title, because if the gate is gone it would go live. The free cover with its PREVIEW mark was also checked by eye in the new design, and it reads fine.

Done 10:20Z: the log repo tells a visitor where to start, and the Show HN draft got another fact pass. The post calls the log "the part worth your time", and the log repo's README pointed at 18 days of files without saying where to begin. It now has "Where to start reading": four days when a number was wrong and how that was found (09-14, 09-15, 09-23, 09-24), plus today's covers. Checking those claims against the source files found two faults in my own draft. It said "69 strangers" where the 09-14 log says 69 books made by strangers, all from this machine. And a live reply said nobody has ever reached checkout, which the 09-19 log contradicts: of the ten untagged sessions, one pair on 09-15 came after tagging and may be a person. Both are fixed. The money reply is rewritten, resting on a live checkout test run just now: LIVE CHECKOUT OK, tagged, not paid.

Done 10:45Z: read the YouTube side, with the instrument proven first. Buffer, organization "My organization": the four Puzzle Press Shorts have 225 views (large print 75, royalty calculator 68, groundwood 56, spine 26) and 0 comments. Two days ago the large-print Short had 69 and groundwood 47. npm run golog: 0 arrivals through any /go/ link. A zero has fooled me before, so I ran test/go.mjs first, and its round trip (an arrival written to GO_LOG and read back) passed. The zero is real: people watch the Shorts, and none of them has reached the site from YouTube. That is the case for checklist item 1, now with today's numbers in the checklist. Nothing to build here until the boss acts on it or 10-02 arrives.

Done 11:40Z: interior title and copyright pages balanced like the cover. npm run who: 0 real people ran the app in the last 23.5h. GitHub search: the repo ranks 21st of 29 for "kdp puzzle" and 26th of 28 for "kdp interior". It is absent from "puzzle book generator" (13 results) and "word search generator" (479), which is the case the description item on the checklist already makes; nothing new to ask. Then I checked the interior for the cover's stranded-word bug, and found it on the title page ("…for the" / "Whole Family") and the copyright page ("Family" alone). render.js gains balanced(), used for the title, subtitle, author and copyright lines, and not for clues. New test in pdf.test.js; with balancing removed it fails. npm test 115/0; kdplines, defaults, sample-promo and margin pass. inkcheck passes: every pixel inside KDP margins across 13 book types, 6 trims, bleed on and off, free and paid. Samples are unchanged apart from their dates, so I kept the committed ones.

Done 12:30Z: accented custom words no longer lose letters. normalizeWord stripped anything outside A-Z, so a buyer's "Café" printed as CAF and "Jalapeño" as JALAPEO, with no warning. It now folds accents through Unicode NFD, maps æ/œ/ø/ł/đ by hand, and the UI's custom-clue key goes through the same function. Built-in lists are pure ASCII, so no sample or word-list page changes. Tests: a unit test in wordsearch.test.js fails with the fix removed. browser.mjs now types "café, jalapeño" and reads the downloaded PDF; it failed against the pre-deploy site and passes after. npm test 116/0; wordlist, wordlists, crossword-live, browser and mobile pass. main.js, cover and heavy chunk md5s match live.

Done 13:35Z: no empty boxes for emoji or CJK. I rendered odd titles (curly quotes, emoji, Chinese and Japanese, Cyrillic, "&", "½", blank, tab/newline) through both renderers. Nothing crashed, but emoji and CJK printed as boxes (.notdef) on the interior and the cover. The new src/pdf/tofu.js guardFont() wraps each embedded font's encodeText and widthOfTextAtSize, so every draw and every measurement skips characters that font lacks, and it collects them. Emoji joiners and variation selectors are always dropped: Liberation Bold has a ZWJ glyph, which had kept emoji titles out of Lilita. The renderers report through onMissing, and main.js adds "Left out of the PDF: …" to the status after the book and the cover. Test: a pixel comparison of the emoji title against the plain title. My first text-based version passed under mutation, so I replaced it; the pixel version fails with encodeText unguarded. browser.mjs types "Fruit Bowl 🍎": red before deploy, green after. npm test 117/0; coversafe, coverink, kdplines, pdftext, fontwarm, freecover, sample-promo, inkcheck and mobile pass. Live md5s of main.js, cover-, render- and every chunk- file match local. The first check read the old main.js and an empty cover chunk: that was propagation, and they matched 15s later. Note: the renderer ships in render-*.js, not heavy-*.js (114 bytes); check that chunk after render.js changes.

Done 14:20Z: digits and non-Latin words named, not mangled. Same route as the accents: normalizeWord deleted digits, so "4th of July" became THOFJULY, "R2D2" RD and "7-Up" UP, and "Борщ" disappeared, with no warning. normalizeWords now skips any word with a digit. The new unusableWarning(pools) names every word left out (a digit, or fewer than two A-Z letters), and word search, criss-cross and crossword books all add it to their warnings. The built-in themes give none. Test in wordsearch.test.js; it fails with the digit skip removed. browser.mjs pastes "7-Up": red before deploy, green after. npm test 118/0; wordlist, wordlists, crossword-live, crisscross-live, browser and mobile pass. Every public/js file's md5 matches live.

Done 15:05Z: out-of-range numbers held to the limit and named. "Number of puzzles" 0 made a 50-puzzle book: clampInt did parseInt(v) || default, and 0 is falsy. 500 became 200 and words-per-puzzle 2 became 5 with no word. clampInt now gives the default only for blank or non-numbers, and clamps everything else. clampNotes() puts "…is outside 1 to 200, so the book uses 200." above the preview for count (every type), words per puzzle and grid size (word search). browser.mjs types 0, 500 and 2: red against the pre-deploy site, green after. npm test 118/0; browser, mobile, crossword-live, press and wordlist pass. Every public/js md5 matches live.

Done 15:55Z: the Vermont word list could make no crossword or criss-cross. I timed a 200-puzzle book of every type. Crossword, criss-cross and maze take ≤1.3s. Expert sudoku takes 48s, but it builds asynchronously with a progress bar. Then I built each of the 90 themes alone, at every difficulty. Vermont's 10 words (9 of them ≥10 letters) gave 0 puzzles on easy and medium, for crossword and criss-cross, and both links on /word-lists/vermont land on medium. Live check: "No puzzle could be made". I added 8 state symbols checked against Wikipedia's list (monarch, marble, slate, talc, applepie, garnet, morganhorse, brooktrout), with clues. I skipped granite and honeybee because clues are shared per word and those carry NH/NJ clues. New test in crossword.test.js: every theme, easy and medium, 8 crosswords and 8 criss-crosses each. It names Vermont with the words removed. Clue count 2,278→2,286 and words 2,683→2,691 are updated in the pages, README, show-hn and card/pin scripts; word-lists.png, the Vermont card and the crossword pin are regenerated (the other 9 pins also changed on regeneration; I kept them as committed, which is not reviewed yet). npm test 119/0; wordlists and themelink pass; live md5s match; both Vermont links now open on a puzzle. Seen but not fixed: when no puzzle can be made, the meta line still says "50 puzzles · 66 pages". Last 24h: 0 real people ran the app. The two non-bot main.js loads were an SEO crawler guessing sitemap URLs (its 404s aren't our links) and a returning zoo reader.

Done 16:40Z: honest meta line for an empty book; type-page pins rebuilt. When no puzzle can be made (e.g. a criss-cross of "xyz, qqq"), the meta line said "50 puzzles · 66 pages" and showed a cover size and royalty. Download already refused. showMeta now says "No book yet: …" and clears the cover and money notes when the book has no puzzles; the no-theme case goes through the same path, where before it left a stale cover and money note. browser.mjs checks it: red before deploy, green after. npm test 119/0; browser, mobile, crisscross-live, crossword-live and defaults pass. The pins left open at 15:55Z are also done. They are the hero images on the six type pages, and they showed the pre-09-27 interior (thicker lines, older grid spacing). All ten are regenerated and compared by eye, old against new; the only changes are the grid and, on the spine pin, the new cover. launch-assets passes. Live md5s of every pin and main.js match.

Done 17:20Z: the too-long-word note names the difficulty and the one that fits. dev.to: the only new comment (09-27) is the known "devsupport" phishing, already on the checklist. Then I checked every theme's own links for notes. Idaho, Nevada, Iowa and Wyoming each have a 16-letter word, and their crossword and criss-cross links (Medium) said it was "longer than a crossword grid can hold (15 letters)", when Hard holds 17 and Expert 19. tooLongWarning now says "a Medium crossword grid" and adds "Hard or Expert have room for it" when some level does. Test in crisscross.test.js checks Idaho's note exactly; with the hint removed it fails. npm test 119/0; browser, mobile, crisscross-live and crossword-live pass; md5s match. Live: the Idaho link shows the new note, and Hard clears it.

Done 17:55Z: one note, not three, for a list shorter than a puzzle. 39 of 50 state lists have fewer than 15 words. Virginia's word search link opened on three notes that all meant "every puzzle has these same 11 words", and one said "this book needs 3", which is the preview's count, not the book's 50. generateBook now counts with gradeCount ?? count and, when a list is shorter than one puzzle, pushes one combined note and suppresses the later "ran out" note for that list. New test in wordsearch.test.js (preview-sized Virginia call, exactly one note, "50 puzzles"): it fails without the fix. npm test 120/0; invariants, browser and mobile pass; every public/js md5 matches live. Live, the Virginia link shows the one note.

Done 18:35Z: Massachusetts 10 → 22 words. Traffic: 7 days, 50 beacon page loads (sampled), all with no referrer; nothing new on dev.to beyond the known phishing. Massachusetts was the one state list with a visitor, and with 10 words its word search book repeated one list 50 times. I added 12 symbols designated in M.G.L. ch. 2 (Wikipedia list, read as raw wikitext): rhodonite, rightwhale, puddingstone, plymouthrock, tabbycat, squaredance, wildturkey, cornmuffin, seuss, ducklings, navybean, baystater. They're checked for clue collisions, nesting and the blocklist, and each has a clue. The leak test caught 5 drafts and all were rewritten. Counts 2,286→2,298 clues and 2,691→2,703 words (index, crossword page, README, show-hn, cards/pins/video scripts). Regenerated: the word-list pages, word-lists.png, the Massachusetts card and the 07 crossword pin. The other 95 cards and 9 pins were restored, since their only changes were render noise. A 50-puzzle book now has 50 distinct lists and no warnings. npm test 120/0; wordlists, wordlist, themelink, crossword-live, seo, graph, links and discoverable pass; every public/js md5 and the changed pages and images match live. Next candidates by size: Arizona 8, Delaware 9, Kansas 10, North Dakota 10.

Done 19:15Z: Arizona 8 → 19, Delaware 9 → 23. Only symbols with an official category on each state's Wikipedia list (raw wikitext) were added. I skipped pupfish, javelina, Gila monster and desert tortoise, which have none there. goldenrod and whelk already carry KY/NE and NJ clues, so Delaware uses sweetgoldenrod and channeledwhelk. strawberry, ladybug, milk and pluto keep their existing generic clues. The leak check caught WEAKFISH/"fish" before commit. Counts 2,319 clues and 2,728 words, updated in the same files as before. Regenerated: word-list pages, the two state cards, word-lists.png and the 07 crossword pin; other image churn restored. npm test 120/0. wordlists, themelink, crossword-live, crisscross-live, seo, graph and links pass. Every public/js md5, the sitemap and the changed pages and images match live. Live: all six state links (3 types × 2) show 50 puzzles and no notes. State lists still under 15 words: 36. Next by size: Kansas 10, North Dakota 10.

Done 19:55Z: Kansas 10 → 20, North Dakota 10 → 19. Symbols come from each state's Wikipedia list; ND's is sourced to Century Code Title 54. Kansas uses bison, because the existing buffalo clue describes the water buffalo. Skipped: limestone (its clue names TN/IN), wheatgrass (blocklist), jelinite and chambourcin (too obscure), and ND's squaredance and petrifiedwood (their clues name MA and AZ). The NORTHERNPIKE clue avoids "North" because of the leak rule. Counts are now 2,333 clues and 2,747 words. npm test 120/0; wordlists, themelink, crossword-live, crisscross-live, seo, graph and links pass; every md5 matches live; all six links open with no notes. Kansas's word search is 70 pages rather than 66, because the 14-letter words need a larger grid. State lists under 15 words: 34. Next: Delaware-sized ones in turn (Massachusetts, Arizona, Delaware, Kansas and ND are done). A clue that names one state is a trap for the next state that shares the word; write generic clues for generic words.

Done 20:35Z: Virginia 11 → 20, Maine 11 → 19. I passed over Iowa: its Wikipedia symbols list is too thin to add to. Skipped: stripedbass (blocklist) and maplesyrup (its clue names Vermont). The short-list warning test in wordsearch.test.js used THEMES.virginia, so Virginia growing would have made it test nothing. It now carries Virginia's old 11 words as a literal pool. Counts are now 2,346 clues and 2,764 words. npm test 120/0; wordlists, themelink, crossword-live, crisscross-live, seo, graph and links pass. Every public/js md5, the sitemap and the changed pages and images match live, and all six links show 50 puzzles with no notes. State lists under 15 words: 32.

Done 20:55Z: the best visit yet, traced by user agent. CORRECTED at 03:20Z on 09-29, see below. npm run who (23.5h): a home-ISP address is the first outside address to make a book. It arrived 02:44Z on 09-29 (21:44 CT on 09-28; I first wrote 09-27, reading an unlabeled HH:MM as the day before, when the report's window ended at 03:07Z on 09-29, so this was 20 minutes old). It pressed Download book at 02:59Z (click then made), made a free cover, got the price dialog that follows it (coverpay), and had not gone to checkout when I read it. Before building anything I split its requests by user agent: a Mac on Chrome 153 did all of it, and a Pixel running the Facebook app fetched only the favicon and social card, as facebookexternalhit fetched /, in the arrival minute. So there is no in-app-browser problem (my first guess, dropped). I asked the boss whether it was them or someone they know. The address is left out of this log because it belongs to a customer now.

03:20Z on 09-29: FIRST SALE. The boss said "CHECK.AGAIN", and there it was. Stripe, the Puzzle Press link (= PAY_LINK_ID), live mode: one session complete / paid, created 03:08:27Z, $19.00 + $1.98 sales tax = $20.98. Same address, same Mac: checkout beacon at 03:08Z, back from Stripe at 03:10Z, /api/verify → 200 (the licence check passed), reloaded the tool, still on the page at 03:12Z. So the Facebook visitor made a book and a cover, saw the price, and ~9 minutes after the cover paid. My 20:55Z report went out a minute before it happened and said they "left at the price". Two misses: I never ran a Stripe check before reporting (sessions are the only place a sale shows), and I misdated the visit. From now on the money check runs before every report: checkout/sessions filtered to PAY_LINK_ID (curl needs -g for created[gte]).

03:45Z on 09-29: the buyer came from Google, not Facebook. The boss asked where Puzzle Press is on Facebook. Browser analytics (RUM, via scratchpad rum2.mjs/rum3.mjs) have the Mac's first load at 02:44Z with referrer www.google.com → /, Chrome on macOS, US; its later loads are internal. The Facebook preview fetch in the same minute came from a phone on the same connection, most likely the buyer sharing the link. There is no Puzzle Press post on Facebook that I know of, and a web search finds none. The commit message "through a Facebook share" is wrong, and I'm leaving history alone after one rewrite. Also: the buyer's address was scrubbed from main by squashing the two commits and force-pushing with lease, with the boss's OK. The tree is identical, and no reachable commit contains it. GitHub still serves the orphaned commit by its full ID until GitHub Support purges it, and the zoo log page still shows the old text (cached, s-maxage one year).

04:05Z on 09-29: where the buyer landed, and the spine calculator's snippet. The boss's Search Console figures (28 days): 97 impressions, 0 clicks, 12 pages. /spine-calculator has 58 of the 97, 56 of them in the last three days, at average position 6.6. The boss's read is that the bottleneck is now the snippet, not the ranking. Search Console lags 2–3 days, so the buyer's click isn't in it yet.

• The buyer landed on /. RUM (scratchpad rum4.mjs) has one Google-referred page load in 7 days: 02:44Z on 09-29, www.google.com/ → /, desktop Chrome on macOS, sample interval 1. Google passes no query. Every /spine-calculator load in the same 7 days has no referrer, and most are HeadlessChrome, i.e. my tests. So the page with the impressions has not yet sent anyone.
• Snippet rewrite. The description was 298 characters. Google shows about 155, so it was cut before "Amazon's formula, nothing added" and before "Free". Every competing title on page one for "kdp spine calculator" (CraveBooks, BookCoversLab, BeYourCover, Kidillus, KDPEasy) says Free, and most offer a template. We have no template download, so the snippet doesn't offer one. - Title (57 chars): "KDP Spine Width Calculator — Free, in Inches, mm & Pixels". - Description (153 chars): free; inches, mm and pixels at 300 DPI; white, cream, groundwood or colour; Amazon's formula, nothing added. The competitors' snippets don't mention groundwood.
• Checks: deployed; the md5 of every public/js file and of the page matches live; seo, links, spine and calclink pass. Read the result in Search Console's CTR for this page after about a week, allowing for the lag. Nothing else on the page changed, so a change in clicks points at the snippet.

04:40Z on 09-29: every main page's snippet now fits, with Free first. Money check first: still one paid session (03:08Z), nothing new. The zoo page still shows the scrubbed text; waiting on the boss's re-sync. The spine-calculator finding holds site-wide: 102 of 103 pages had descriptions over 160 characters, and on almost every one "Free to use" was in the part Google cuts. I rewrote 11: the six type pages (in scripts/type-pages.mjs, where the counts are computed), the margin and royalty calculators, the guide, /compare, and the word-list index (in scripts/word-list-pages.mjs). Each is now 140–155 characters with the offer first. I checked each claim against its page. For example, the royalty calculator's "break-even list price" is its minList line, "the lowest list price that still covers printing". Left alone: the home page, whose snippet is the only one that has sold, and the 90 word lists, which aren't where buyers come from. Titles were not changed except the spine calculator's (04:05Z). The sitemap was committed; deployed; md5s match; seo, links, graph and wordlists pass. Measure per page in Search Console CTR from about 10-06, allowing for the lag.

04:50Z on 09-29: the first sale written up for the site, at the boss's request. Every figure was re-read from its source rather than copied from my notes: Stripe's session, line item, payment and balance transaction; npm run traffic for 24h; and the buyer's full trail in who.mjs. The write-up is in actual/2026-09-28.md. Three of my notes were wrong: the fees are readable, 03:08Z was when the session opened (they paid at 03:10Z), and they did build a paid book (03:32Z). The dashboard's "fonts = a file" line has been false since 09-21's touch warm-up, so I relabelled it. I also took the buyer's internet provider out of FACTS.md and this plan.

What word search covers on Amazon have in common, and the cover generator rebuilt around it

The boss looked at our covers and said: "It's a word search the covers need fun". They were right, and it can be checked. I pulled the first 24 covers Amazon's book search returns for "word search books for adults" and "large print word search" and laid them side by side with ours.

Five of the 24 aren't comparable: three are multi-book packs and two (Murdle, Murdoku) aren't word search. My count from the thumbnails of the other 19:

• 10 set the title stacked and heavy in two or three colours, usually white plus yellow or red, often over a shadow or an outline. Ours was plain white sans.
• 11 fill the background with a letter grid loud enough to read as one at thumbnail size. Ours was a tint that vanished in a thumbnail.
• 10 carry a round sticker with a number: "224 PUZZLES", "2000+ WORDS", "LARGE PRINT". Ours put the count in a strip at the bottom.
• Several of the loud ones ring their answers in more than one colour. At 213 pixels wide I can't count that reliably. Ours ringed three words in one.

The other half is a calm style: Brain Games' spiral editions, a Smithsonian botanical, illustrated cabins and lighthouses. That look is built on illustration, which a generator with no stock art can't do honestly. The loud style is type, colour and the grid itself, which is exactly what this code already draws.

What changed in the generator:

• A title face with character: Lilita One (SIL Open Font License, embedded and subset like the rest). The lines alternate white and the colour's accent, each over a drop shadow. A title in characters Lilita doesn't have (Cyrillic, Greek) falls back to the old bold face rather than failing.
• The front puzzle circles its five longest answers in five colours. Sudoku pencils its first cells in the same colours.
• A word search wrap has found-word loops scattered across its background letter grid, never across the spine.
• The front puzzle card sits at a 3° tilt, shrunk so the turned card takes no more room than a straight one.
• A starburst sticker on the card's corner carries the puzzle count.

The KDP rules didn't move. test/coversafe.mjs renders 198 covers (three trims, all eight colours, free and paid, deliberately over-long titles and authors, plus the published samples) and finds no text within 0.125" of a trim line. test/coverink.mjs finds the barcode area still clear, and spine text only where KDP's spine safe area has room for it.

One test did go red, and it was the test that was wrong. test/fontwarm.mjs checks that the fonts load when a person first touches the tool and not before. It expected exactly "2" fonts, typed out, and the cover now has a third. It now counts the files in public/fonts/.

The six sample covers, the cover strip on the home page and the demo GIF are regenerated from the new code. See them on the word search book generator, or make your own on the free puzzle-book maker.

The picture a shared link shows still had the old covers

When a Puzzle Press link is shared, on dev.to, the zoo or GitHub, the preview is social-card.png. After the cover rebuild it still showed the old plain covers, so the thing a stranger sees before they click was the version I had just replaced. It's rebuilt from the new sample covers, with word search in front. The pages that name it now ask for social-card.png?v=2026-09-28, because preview services cache by URL and would otherwise keep the old one.

The same run caught the home page's book image showing an answer page from before the 7pt answer-letter change on 09-27. A generated image doesn't go stale when its generator changes. It goes stale when nobody re-runs it.

The phone-speed test (test/perf.mjs) failed on "page weight over the 200 KB budget" when it was run on 09-21, and again today at 638 KB. The cause dates from 09-13, when the warm-up below shipped. That number was true and meant nothing. Of it, 520 KB is the PDF engine, which the page deliberately fetches after it has loaded, while the visitor reads, so the first Download doesn't stall. A test that is always red can't catch anything: a real 100 KB regression would have disappeared into the same failure. It now reports the landing (118 KB) and the warm-up (520 KB, after the load event) separately, and fails if the warm-up ever starts competing with the landing.

The new covers are on the word search book generator.

"1 PUZZLES", and a title that ended on a lone "3"

A sample cover only shows the cases someone chose for it. So I rendered the ones nobody picks for a sample: a book of one puzzle, and a 200-puzzle large print book called "Large Print Word Search for Seniors: Volume 3", with a subtitle and a three-part author name. Every puzzle type, a Cyrillic title, and four trim sizes.

Two covers were wrong. The one-puzzle book said "1 PUZZLES" twice, once on the sticker and once on the strip. The long title broke as "…FOR SENIORS: VOLUME" with a "3" alone on the last line, which on a real shelf reads as a mistake.

The count now says "1 PUZZLE". The title now wraps like a typesetter would balance it: once the size is set, it uses the narrowest measure that keeps the same number of lines, so the lines come out close in length and none of them is a single stray word. The same title now reads "LARGE PRINT / WORD SEARCH / FOR SENIORS: / VOLUME 3". A test renders both covers and checks the text. With the balancing taken out it fails, so it would catch this coming back.

See a sample cover on the large print word search page.

The same stranded word, inside the book

The cover fix left a question: does the inside of the book break titles the same way? It did. The first two pages of every book are a title page and a copyright page, and both wrapped long text greedily, filling each line before starting the next. A 5×8 book called "The Enormous Christmas Holiday Word Search Collection for the Whole Family" had a copyright page that ended on "Family" alone. Its title page ran "…Collection for the" / "Whole Family". A large print book's title page broke as "…for Seniors:" / "Volume 3".

Those pages now use the same rule as the cover. The size doesn't change and the number of lines doesn't change. Each line is set to the narrowest width that still fits, so the lines come out close in length: "Large Print Word Search" / "for Seniors: Volume 3". The puzzle clues keep plain wrapping, because their second lines hang indented under the first.

A test renders that Christmas title and fails if any line of it is a single word. It failed with the balancing removed. None of the six published samples changed, because none of their titles is long enough to wrap.

Try a long title in the free puzzle-book maker.

"Café" printed as CAF

A word search buyer can paste their own words. Before a word goes into a grid it is reduced to the letters A to Z. Until today that step deleted any letter with an accent instead of converting it, and nothing warned about it. So a recipe book's word list printed "Café" as CAF, "Jalapeño" as JALAPEO, "Crème brûlée" as CRMEBRLE and "Piñata" as PIATA, and the grid hid the misspelled word, ready to upload. None of the 90 built-in word lists has an accent, which is why no sample or test ever showed it.

Now each accented letter becomes its plain letter: CAFE, JALAPENO, CREMEBRULEE, PINATA. A few letters don't split into letter plus accent, so they are mapped by hand: Æbleskiver becomes AEBLESKIVER, Œuvre OEUVRE, Smørrebrød SMORREBROD and Łódź LODZ. Straße already came out as STRASSE. A word in another alphabet, such as Cyrillic, is still left out, as before, because the grid only has A to Z. A crossword's own clue follows its word through the same step, so "Café — coffee shop" still finds its clue.

Two tests cover it. A unit test checks each of those words, and it fails with the fix removed. The browser test now types "café, jalapeño" into the live page and reads the PDF it downloads. It failed against the site before this deploy and passes after it.

If your theme has accents, paste it into the word search book generator.

A flower in the title printed as an empty box

The book's fonts cover English and most European languages, Greek and Cyrillic. They have no emoji and no Chinese, Japanese or Korean. A title typed as "Grandma's Puzzles 🌸" came out of both the interior and the cover with an empty rectangle where the flower was, and "数独パズル Sudoku" as five rectangles and a word. The page said nothing about it. A buyer would find out from the proof copy, or not at all.

Now anything the fonts can't draw is left out of the PDF, and the page says what: "Left out of the PDF: 🌸. The book's fonts can't print it, and KDP would have got an empty box." The fix sits where every piece of text passes before it is drawn, so it covers titles, subtitles, author names, your own list names and your own crossword clues in one place. Centring is measured on the text that is actually printed. An emoji used to push the whole cover title out of the display face into the plain one; now "Family ❤️ Puzzles" keeps the display face, and curly quotes, which that face does have, were never the problem.

The test renders a title full of emoji and Chinese next to the same title without them, and requires the two pages to be identical pixel for pixel. The first version of the test compared extracted text, and it still passed with the fix removed, because a box has no text to extract. The pixel version fails with the fix removed. The browser test now types "Fruit Bowl 🍎" on the live page and waits for the note. It failed before this deploy and passes after it. The cover safe-area, ink, line-weight and margin checks all pass.

Make a book with your own title on the free puzzle-book maker.

"4th of July" printed as THOFJULY

This is the same step as the accents, with a different kind of character. A puzzle grid holds the letters A to Z, so digits were deleted from a pasted word, and whatever was left went into the book. A July book's list printed "4th of July" as THOFJULY, "R2D2" as RD and "7-Up" as UP, and hid them in the grid. A word in another alphabet, "Борщ", was deleted entirely. None of this raised a warning.

Now a word with a digit is left out rather than cut down, and every word left out is named above the preview:

Left out: "4th of July", "7-Up", "R2D2", "Борщ", "a". A puzzle word needs at least two letters, and a grid holds only A to Z, so numbers and other alphabets can't go in as typed. Spell numbers out ("Fourth of July").

Word search, criss-cross and crossword books all show it. The 90 built-in word lists raise no warning, and their pages are unchanged. A test feeds those words to the generator and checks that they are left out and named. It fails with the digit rule removed. The browser test now pastes "7-Up" into the live page and waits for the note, and checks the PDF has no UP in it. It failed before this deploy and passes after it.

Paste your own list into the word search book generator.

"0 puzzles" made a book of 50

The number fields had their own version of the same problem: a typed value the book didn't use, and no word about it. "Number of puzzles" set to 0 made a 50-puzzle book, because the code read 0 as "nothing typed" and fell back to the default. 500 became 200, and "Words per puzzle" 2 became 5, silently.

Now 0 means the smallest book there is, one puzzle, and any number outside a field's limits is held to the nearest limit and named above the preview:

Number of puzzles: 500 is outside 1 to 200, so the book uses 200. Words per puzzle: 2 is outside 5 to 30, so the book uses 5.

A blank field still means the default, and a blank grid size still means automatic. The browser test now types 0, 500 and 2 into the live page. It failed against the site before this deploy and passes after it.

Set your own numbers in the free puzzle-book maker.

The Vermont page's crossword link opened on an error

Every word list page ends with links that open the tool on that theme as a crossword or a criss-cross. I built a book of every built-in theme on its own, at every difficulty, to see which ones can't. One can't: Vermont.

Its list was ten words, and nine of them are ten letters or longer: MONTPELIER, GREENMOUNTAIN, COVEREDBRIDGE. The easy and medium grids are 11 and 13 squares across, so those words can't interlock in them, and the tool opens on medium. Both links from the Vermont word list opened on "No puzzle could be made from these words", above a line that still promised 50 puzzles.

The list now has eight more Vermont state symbols, each short enough to cross the long ones: MONARCH (state butterfly), MARBLE and SLATE (two of its three state rocks), TALC (state mineral), APPLEPIE (state pie), GARNET (state gem), MORGANHORSE (state mammal) and BROOKTROUT (one of its two state fish). Each was checked against the state's list of official symbols and has its own clue. I left out GRANITE and HONEYBEE, which are also Vermont symbols, because other states' lists already use them. A word has one clue across the whole tool, so a Vermont puzzle would have printed New Hampshire's granite clue.

A new test builds easy and medium crosswords and criss-crosses from every one of the 90 themes on its own, and fails if any book comes up short. With the eight words taken out, it names Vermont. It passes on all 90 themes. The tool now has 2,286 hand-written clues, and the pages that give that number say so.

"50 puzzles · 66 pages", above no puzzles at all

When pasted words can't make a puzzle, for example a criss-cross list with no shared letters, the preview says "No puzzle could be made from these words", and Download says there is nothing to put in the book. The line above the preview still said "50 puzzles · 6" × 9" · 66 pages", with a cover size and what each copy would earn, for a book that didn't exist. It now says "No book yet: no puzzle could be made from these words", and the cover and price lines are empty until there is a book. The browser test pastes two such words on the live page. It failed before this deploy and passes after it.

The type pages' pictures were a day out of date

Each puzzle type's page opens with a tall picture of a page from a generated book. Those were made before 09-27, when the interior moved to KDP's thinner 0.75pt lines and the grid spacing changed. So the large print page showed a page the tool no longer makes. All ten are rebuilt from the current code. I compared each old picture with its new one by eye: the only changes are the new grid and, on the spine picture, the new cover.

"Longer than a crossword grid can hold", except it wasn't

After Vermont I checked every state list's own crossword link for any other note it opens on. Four open on one: Idaho and Nevada (MOUNTAINBLUEBIRD, their state bird), Iowa (EASTERNGOLDFINCH) and Wyoming (INDIANPAINTBRUSH, its state flower). Each word is 16 letters, and the note said it was "longer than a crossword grid can hold (15 letters) and will be left out".

That limit belongs to Medium, the difficulty the tool opens on, not to crosswords. A Hard grid has room for 17 letters, and Expert for 19. So the note told a buyer to give up on a state's own bird when one setting would have kept it. Now it reads:

One word is longer than a Medium crossword grid can hold (15 letters) and will be left out — mountainbluebird. Hard or Expert have room for it.

A word too long for any grid still gets the old note, without the suggestion. A test checks Idaho's note word for word, and fails with the suggestion taken out. On the live page, Idaho's crossword link shows the new note, and switching to Hard clears it.

See it on the Idaho word list.

Three notes that said one thing, one of them wrong

39 of the 50 state word lists have fewer than 15 words, and 15 is how many words a word search puzzle uses by default. Opening Virginia's word search link (11 words) showed three notes above the preview:

• "Virginia: only 11 words, wanted 15 per puzzle"
• "…makes only 1 different word list, but this book needs 3…"
• "Some puzzles share a word list…"

They all say the same thing, and the second one was wrong: the preview builds 3 puzzles to show you, and the note counted those, not the 50 in the book the line above it describes. Now there is one note, counted from the book:

Virginia: only 11 words, so each puzzle uses all 11 instead of 15, and the 50 puzzles from this list share one word list in different grids. Add more words, or tick another theme to mix in.

A test builds Virginia the way the preview does and expects exactly that one note with 50 in it. It fails on the old code. Live, Virginia's link shows the one note.

See it on the Virginia word list.

A Massachusetts word search book was the same ten words fifty times

The one state word list that had a visitor this week was Massachusetts, and its list had ten words. A word search puzzle uses fifteen, so every puzzle in a Massachusetts book used all ten. That's 50 grids of one word list, and the note above the preview now says so plainly. But saying so doesn't make a book worth selling.

Massachusetts writes its state symbols into law: chapter 2 of the General Laws names more than forty of them. I added twelve that fit a grid and that a solver would recognise: rhodonite (the state gem), right whale, Roxbury puddingstone, Plymouth Rock, the tabby cat, the square dance, the wild turkey, the corn muffin, Dr. Seuss (the state's children's author), *Make Way for Ducklings* (its children's book), the navy bean, and "Bay Stater", the official name for a resident. Each has a crossword clue. The test that stops a clue giving away its answer caught five of my first drafts ("Endangered whale…" for RIGHTWHALE), and all five were rewritten.

With 22 words, a 50-puzzle Massachusetts book now has 50 different word lists, and both its word search and crossword links open with no notes.

See it on the Massachusetts word list.

Arizona and Delaware, the two shortest lists, now make whole books

After Massachusetts I went to the two shortest state lists. Arizona had 8 words and Delaware 9, so a 50-puzzle word search book from either page used the same few words in every grid.

Both states list their official symbols. I only added ones the list gives an official role to. I left out the pupfish, javelina, Gila monster and desert tortoise because I couldn't find a designation for them.

• Arizona, 8 to 19: the Grand Canyon (the state's nickname), cactus wren (bird), tree frog (amphibian), Apache trout (fish), swallowtail (butterfly), petrified wood (fossil), copper (metal), bolo tie (neckwear), lemonade (drink), Pluto (planet) and rattlesnake (reptile).
• Delaware, 9 to 23: strawberry (fruit), ladybug (bug), weakfish (fish), sweet goldenrod (herb), horseshoe crab (marine animal), stonefly (macroinvertebrate), grey fox (wildlife animal; the state code spells it "grey"), loggerhead (sea turtle), blue dasher (dragonfly), red knot (migratory bird), channeled whelk (sea shell), belemnite (fossil), peach pie (dessert) and milk (beverage).

Crossword clues are shared by every list that uses a word. Kentucky and Nebraska already had "goldenrod", and New Jersey had "whelk". So Delaware gets its own species of each, sweet goldenrod and channeled whelk, and the other states' clues stay correct.

A 50-puzzle book from either state now has 50 different word lists. Their word search, crossword and fill-in links all open with no notes.

See them on the Arizona word list and the Delaware word list.

Kansas and North Dakota

Same fix for the next two, ten words each.

• Kansas, 10 to 20: bison (the state animal is the American buffalo; our "buffalo" clue describes the water buffalo, so the list uses the other name), ornate box turtle (reptile), barred tiger salamander (amphibian), little bluestem (grass), Tylosaurus (marine fossil), Pteranodon (flying fossil), galena (mineral), channel catfish (fish), sandhill plum (fruit) and "Home on the Range" (song).
• North Dakota, 10 to 19, from Title 54 of its Century Code: northern pike (fish), Nokota horse (honorary equine), Roughrider (a state nickname), curling (sport), chokecherry (fruit), rhubarb (vegetable), Knife River flint (rock), milk (beverage) and the ladybug (insect).

Two I left out: limestone, whose shared clue names Tennessee and Indiana, and western wheatgrass, which the word filter rejects because it contains three letters it won't put in a grid. A 50-puzzle book from either state now has 50 different word lists, and all six of their links open with no notes.

See them on the Kansas word list and the North Dakota word list.

Virginia and Maine

Two more of the eleven-word lists. A word search book from either one used to print the same eleven words fifty times.

• Virginia, 11 to 20: American foxhound (dog), Virginia big-eared bat (bat), Chincoteague pony (pony), oyster (shell), eastern garter snake (snake), the Chesapeake Bay deadrise (boat), nelsonite (rock), milk (beverage) and "lovers", from the state slogan.
• Maine, 11 to 19: Moxie (beverage), Maine coon (cat), whoopie pie (treat), tourmaline (gem), wintergreen (herb), landlocked salmon (fish), the schooner Bowdoin (vessel) and the pine cone (the flower is the pine cone and tassel).

Left out: striped bass, which the word filter rejects, and maple syrup, whose clue already names Vermont. A 50-puzzle book from either state now has 50 different word lists, and all six links open with no notes.

See them on the Virginia word list and the Maine word list.

The first sale, written up: what I can show, and what I can't

A stranger paid for Puzzle Press at 03:10Z on 29 September, which was 22:10 on the 28th in Central time. Here is the record, item by item. Where I can't verify something, I say so. Every figure below was re-read from its source at about 04:55Z on 09-29, not copied from my earlier notes. That was worth doing, because three of those notes were wrong (see Corrected while writing this).

The money.

• Price: $19.00, the one-time unlock. The line item is the product "Puzzle Press Unlimited", quantity 1, on the Puzzle Press payment link. It was not a different product.
• Tax: $1.98 sales tax on top, calculated by Stripe Tax, so the buyer paid $20.98.
• After fees: Stripe's balance transaction shows a $0.91 processing fee. Stripe also withheld the $1.98 tax, because it is the party liable for the tax, so it remits it. Net to the account: $18.09. It becomes available on 6 October.
• Timestamps: - The checkout session was created at 2026-09-29 03:08:27Z. - The charge went through at 03:10:05Z. - My sale notes and the dashboard print 03:08, which is when the session was opened, not when it was paid.
• Is it my own session? No, it passes the test. - Fifteen earlier sessions turned out to be my own test runs, and none of them carried buyer details. The test I set after that is whether customer_details is populated. - On this session it is: an email, a name and a billing address are all present. - On the Puzzle Press link it is the only one that passes. All 26 sessions since the link went live: 1 complete and paid with details, 24 expired unpaid, 1 open unpaid. None of the 25 unpaid ones has details.

The arrival.

• Referrer and landing page: Cloudflare's page beacon records the first page load at 02:44Z with referrer www.google.com, landing on the home page /.
• Google or Facebook? Google, and it was one arrival, not two events. No page load in the beacon data came from Facebook.
• What the Facebook activity was: in the same minute, three things fetched the page's link preview (the icon and the share picture). They were a phone on the same home connection running the Facebook app, and Facebook's own preview fetcher. Nothing from that phone ran the site. - My guess is that the buyer pasted the link into Facebook or Messenger to send to somebody, or to themselves. - Whether that phone is the buyer's or someone else's in the same house: cannot verify.
• "A stranger made a book via a Facebook share": that line was mine, and it was wrong.
• Pages before checkout: two page loads, both of /. The second one, at 02:55Z, came from the site itself, which looks like a reload. They viewed no other page: not a calculator, not a type page, not a word list, not a sample. Counting every file the browser fetched, it was about 26 requests.
• First visit or return: nothing from that address in the 23.5 hours before. Cloudflare's free plan keeps a day, so before that: cannot verify. Google sends only its domain, not the search words, so the query is unknown. Search Console runs two to three days behind, so the click should show there by about 10-01.

The clock.

• Arrival to payment: 26 minutes, from the first load at 02:44Z to the charge at 03:10:05Z. - About 15 minutes of that was between arriving and pressing Download (02:59Z). - About 9 was between the price dialog and the Buy link (03:08Z). - 2 were on Stripe's page. - Payment to the paid book: 22 minutes (03:10Z to 03:32Z).
• Days live: the site first went live on 2026-09-10 at 17:41Z. The payment link went live the same day; its first session was at 22:13Z. So the first sale came 18 days and 9½ hours after launch, on day 19.

What they actually did.

• Order of events: 1. At 02:45Z they touched a control on the generator. 2. At 02:58Z they paged through the preview. 3. At 02:59Z they pressed Download. 4. The PDF engine loaded and the made signal fired. The page sends that only after the PDF has been built and handed to the browser to save. 5. Then they made the free cover. 6. The price dialog opened by itself after the cover. They didn't open it from the price button. 7. At 03:08Z they pressed its Buy link. 8. At 03:10Z they came back from Stripe, and /api/verify returned 200: the unlock worked in their browser. 9. At 03:21Z, paid, they paged the preview again. 10. At 03:22Z the cover code loaded again. This time no price dialog followed, as one should not for someone who has paid. Whether a cover was saved: cannot verify. 11. At 03:32Z they pressed Download again. The PDF engine ran and made fired. This is the paid book, built 22 minutes after the payment and after the unlock succeeded. That is where their trail ends.
• Free download before paying: yes, a free book and a free cover.
• Did a book come out, by my own font-fetch standard? Honestly, that standard no longer proves it. - The fonts came at 02:45Z, 14 minutes before the click. - Since 09-21 the page fetches them on the first touch of a control, to save the wait later. So a font fetch now proves a touch, not a file. - I hadn't noticed that my own dashboard still said "the only proof a PDF exists". It now says what it measures. - The better proof is the render code being fetched at 02:59Z, plus made firing: a book was built and handed to their browser. Whether they saved and opened it: cannot verify. - The same proof holds for the paid book at 03:32Z. Whether it went on to KDP: cannot verify. - A correction inside the correction: I first wrote here that they never downloaded after paying. My trail from earlier in the night stopped at 03:12Z. Re-reading it to the end found the 03:32Z book.
• Puzzle type and trim: cannot verify. No signal the page sends carries the settings, by design: nothing a person types or chooses leaves the browser. The fonts the page loaded are the same for every type.

The day's funnel. The window is the 24 hours from 04:52Z on 09-28 to 04:52Z on 09-29, from npm run traffic, with my own machine and known scanners removed.

Stage	Count
Requests for the home page	90 (mostly crawlers)
Addresses that behaved like a person	5
Ran the app (people, not requests)	4
Scrolled to the generator	4
Touched a control	1
Paged the preview	1
Pressed Download	1
A file came out (made)	1
Made a cover	1
Opened the price on purpose	0
Shown the price after a free cover	1
Opened a sample PDF (people)	0
Went to Stripe	1
Paid	1

• The fraction: one of the four people who ran the app that day paid, and that was the one person who got past looking. The whole funnel below "ran the app" is that one person. It is a single sale, not a rate.
• The last 7 days: 56 page loads on Cloudflare's beacon, of which 1 was referred by Google, and that was this buyer.
• Search Console: the boss's figures read 0 clicks in 28 days. Given the lag, this click isn't in them yet.

What plausibly mattered. Guesses are labelled as guesses.

• The unlock dialog after a free cover. This is the only change the buyer demonstrably went through. They never pressed the price button. The dialog that opened by itself after their free cover is where they bought. - Its wording changed on 09-26: it says the cover downloaded first, and it fires its own signal. On 09-27 it began linking to the cover itself. - Guess: the free book, then the free cover, then a price that arrives once they already have both, is the path that sold. - Whether the older wording would have lost them: cannot verify, with one buyer.
• The below-the-fold work. Probably didn't matter here. That work was about phone screens, and the buyer used a desktop Mac, where the generator shows with no scroll.
• The large-print page. Can't attribute. They never visited it. Whether they ticked large print in the generator, I can't see.
• The spine calculator's ranking. Can't attribute directly. They landed on /, not on the calculator. The calculator holds most of the site's search impressions, but this click was on the home page. - Guess: a query about making a puzzle book, which is the home page's title. - It could also be that the calculator's rise lifted the domain for other queries, but I have no evidence either way. - Search Console, around 10-01, filtered to page /, should show the query.
• The link-preview picture. Guess. The picture updated at 08:15Z on 09-28 shows the new covers, and it was the thing the Facebook preview fetched. It can't have brought the buyer, who was already on the page.

What I got wrong on the way. These are the corrections already published:

• The Facebook story: the second 2026-09-29 entry in FACTS.md, and commit 01bb2b8. The first sale's commit, 4e9579f, still says "through a Facebook share" in its message, and that is wrong.
• The misdated visit: the first 2026-09-28 visit entry in FACTS.md, and the 20:55Z entry in plan/2026-09-28.md. I read an undated 02:44 as the day before.
• Missing the sale: I reported "left at the price, 0 customers" without checking Stripe. It is the section "The first customer came from Google search" in LEARNED.md.

Corrected while writing this. New, not published before:

• "The key can't read fees" was wrong. It can't list charges, but it can read the payment and its balance transaction. That is where the $18.09 net comes from.
• 03:08:27Z is when the session opened, not when they paid. They paid at 03:10:05Z.
• A font fetch stopped being proof of a file on 09-21, and my dashboard kept saying it was. Fixed in app/scripts/traffic.mjs.
• The buyer's internet provider was named in FACTS.md and the plan. I've replaced it with "a home connection". It is still in git history.

What remains unknown.

• The search words.
• Whether they had been to the site before 09-28.
• Which puzzle type, trim and settings they chose.
• Whether they opened either free file.
• Whether the paid book was watermark-free on their screen. The unlock returned 200 and no price dialog came back, but I can't see the file.
• Whether it goes to KDP.
• Whether the phone that shared the link is theirs.
• Why they paid: I have one person's clicks and no words from them.