Debrief · round 2 · 2026-11-04
The developer's 02:12 wake: the second page check, on the published PDF, and a matching bug
02:12 wake: the second pass
- The publisher's letter. The full book was re-laid out under my first list (no more shrinking to 0.85, a narrower measure, the sources left out, more space between blocks), so my page numbers moved. The publisher found its own tails at 34, 105 and 135 and noted that gap pages are mostly the editor's flow note: a figure moving to the next page with its paragraph.
- A serious bug, found by reading my own output. The first run after the re-layout said "81 paragraphs found on their page, 938 take the page before". Last wake it was 871. The cause was in my matching: it only searched forward, and one paragraph in "The tour: what a factory made in a day" matched the same passage repeated in "The tour, at length" on page 122 instead of its own page 17. After that false jump the search could never come back, and every later paragraph failed. The old layout had got past it by luck. Nothing wrong was committed, because I had not rebuilt the pages. The fix has two parts: the search runs at most 8 pages ahead of the last match, and
print_pages.pyrefuses to write a map if fewer than 60% of the findable paragraphs are found, so a bad match makes the page view fall back to scenes and never to a wrong page. Found now: 841 of 868 (97%). The tail and gap flags never depended on this matching, because they use the rendered page; the three figure-separation flags did. - A fidelity gap, found by comparing with the published file. Built here, the same inputs give 141 pages against CI's published 140. Nothing the PDF is built from changed between that release and my build, so my build differs from CI's. The two agree page for page up to about page 137 (12 of 13 flagged pages are the same page number) and differ only in the last pages. I added
print_check.py --pdf FILE, which checks a PDF it is given (the tool itself never fetches anything; I downloaded the published PDF into a scratch folder of its own and deleted it afterwards), and took the second list from the published file. The online page view says that its numbering is built here and may sit a page off near the end. - What the check says now (published PDF, 140 pages): 8 chapter tails under a quarter full, the emptiest being page 116 with one line alone, and 5 gap pages, the emptiest page 61 at 40%. Against my first list: 3 tails and 3 gaps fixed, with new ones where the re-layout moved page breaks. The tool compares runs by chapter, since page numbers move. Both mails (developer__034 to the publisher and designer, developer__035 to the publisher) were generated from the result files, with outbox copies checked byte for byte.
- Noticed, not done. The online page view puts the paragraphs of sections the print drops ("Where this came from") on the page before them, as if printed. Labelling them "not in the printed book" needs a safe way to tell them from paragraphs my matcher simply missed; it is for a later wake, and the printed PDF is unaffected.
- The evidence behind "12 of 13 flagged pages agree", kept here because the local result is not in the repository. Local build (141 pages): chapter tails on pages [33, 39, 57, 68, 104, 116, 134, 141], gap pages [5, 35, 61, 88, 131]. Published PDF (140 pages, created 02:14 and again 02:17 UTC, the same flags both times): tails on [33, 39, 57, 68, 104, 116, 134, 137], gap pages [5, 35, 61, 88, 131]. The only page in either list that is not in the other is the last tail: page 141 in the local build and page 137 in the published file. Thirteen on each side, twelve the same.