An EPUB page break has at least three observable forms: a marker saved in the manuscript, a boundary in generated XHTML, and a visual transition in a reading system. Searching the output for the literal words “page-break” is not enough. An exporter may realize the break by splitting content into another XHTML file without leaving that token behind.
STUDIO-272 has two evidence runs. An earlier synthetic export saved one manuscript marker but returned zero literal page-break tokens in generated XHTML, so that search alone could not settle the question. A later fixed specimen placed one marker before line three, where the sentinel MIDWAY-ANCHOR-272 begins. After export, preceding text was in p001.xhtml, the sentinel was in p002.xhtml, and the source marker leaked into neither file. The resulting EPUB has SHA-256 978f8b1e127f97c3a8112d323009940394e18425adf9f0056fadfdf35553a66a. Physical-reader pagination remains unmeasured.
Give the Break a Visible Sentence on Each Side
Use a short source whose boundary can be recognized without line-number guesswork. The retained specimen names line three MIDWAY-ANCHOR-272 and keeps a known sentence immediately before it. Record the source path and digest before insertion. These anchors let an inspector identify both sides even after the exporter removes its private notation.
Avoid placing the first test at a chapter start, where a natural chapter boundary could hide a failed marker. A mid-chapter line makes the expected transformation distinct: content before the marker stays in one generated document, while the sentinel begins the next.
Preview Line Three Before Applying the Marker
The stored sequence uses the page-break command in dry-run mode before applying it. The target is line three of the synthetic source. Review the preview, returned selection, and proposed change before allowing the write. A successful preview describes a candidate operation; it does not prove that the marker reached disk.
Apply the same command only after the target is correct. Preserve the response and reread the manuscript as a separate step. If the marker lands above or below MIDWAY-ANCHOR-272, stop and keep that failed source for diagnosis rather than moving it by hand without a record.
Confirm One Marker in the Saved Manuscript
Reinspection of the fixed source found one saved 《《｜Page Break｜》》 marker. The sentinel remained after it, and the surrounding text was unchanged. The acceptance tuple at this layer is therefore one target line, one marker, and one intact anchor. A command response without this saved readback is incomplete evidence.
This check also catches accidental duplication. Two consecutive markers could still lead to a package that opens, but they no longer represent the planned single boundary. Compare the saved file digest with the command record before export so the generated book can be tied to the exact manuscript inspected.
Inspect the p001-to-p002 Split
Export the saved source and open the EPUB as a ZIP. In the measured package, the sentence before the marker occurs in p001.xhtml. MIDWAY-ANCHOR-272 occurs in p002.xhtml. Searching the generated XHTML finds zero leaked source markers. This file transition is the structural evidence that the later run lacked when it searched only for a literal CSS-like term.
The EPUB digest is 978f8b1e127f97c3a8112d323009940394e18425adf9f0056fadfdf35553a66a. Keep that digest beside the extracted p001 and p002 files. A later export can reuse a filename, so its path alone does not identify the result.
Understand Why Zero Literal Tokens Is Not a Failure
The earlier run found one source marker and zero page-break strings in XHTML. That result is genuinely inconclusive. It neither proves a broken export nor proves a working visual break. The relevant output might be a file boundary, an element class with another name, or a stylesheet declaration outside the searched text.
Recovery begins by locating the known sentences on both sides, examining generated file boundaries, and reading the stylesheet. If the sentinel stays in the same XHTML document and no equivalent break structure exists, reject that export. If it starts a new document, record the split rather than inventing a token count.
The related baseline run also inserted ruby and produced one ruby element after export. Its package checks reported a valid book with navigation present. A page-break correction must not silently remove that independent formatting result or change unrelated navigation. Rerun previously accepted inspections on every new digest.
Do not patch p001.xhtml and p002.xhtml inside the archive to manufacture the desired split. Fix the manuscript marker or export selection, create a differently named EPUB, and reopen it. Source-level recovery is reproducible; a one-off ZIP edit is not.
Leave Physical Pagination as an Open Test
The p001/p002 split proves a generated document boundary for the fixed artifact. It does not show whether every reader starts the sentinel on a new screen, page, or column. Font size, writing mode, viewport, and pagination engine can affect presentation. No named physical device was operated for this result.
A later visual test should record reader name and version, device, orientation, text size, writing mode, EPUB digest, the last visible sentence before the transition, and the first visible sentence after it. Until that receipt exists, say that the structural split was measured and the displayed break remains unmeasured.
Accept One Marker, One Split, and No Leakage
The final handoff identifies the original source digest, the dry-run and apply commands, line three, one saved marker, the intact sentinel, the EPUB digest, p001.xhtml, p002.xhtml, and zero leaked notation. Acceptance requires the preceding sentence in p001 and MIDWAY-ANCHOR-272 in p002. A second operator can reproduce that decision by reopening the retained files.
Keep the older zero-token observation as a lesson, not as a contradictory success or failure. It shows why a literal search was too narrow. The later boundary artifact supplies the missing structural evidence, while reader appearance remains explicitly open. That layered conclusion is more useful than a generic statement that the EPUB “looked correct.”