docs: correct three source claims in the ADR 0018 review notes

All three failed the same way: taken from the PDF's extracted text layer rather
than from rendered pages. In this package the text layer misplaces checkbox and
Y/N columns -- marks collapse upward toward the first rows of a table -- so a
filled form reads as a mostly-empty one.

- Page count: 73, not 24.
- Pump checklist: 3 of 20 items blank (6, 12, 20), not 14 of 20. Rotation
  direction and power consumption are both marked and initialled. Independently
  confirmable from the extraction: the block carries 17 X marks and 17 sets of
  initials against what the text layer shows as six items.
- The orphan failure is the line confirming automatic recovery from simulated
  power-loss events, not a PID-tuning line. The PID line is marked pass; its
  note describes a defect plus a workaround, which makes it an instance of open
  question 6 (pass with exception), not question 7.

The third correction changes a recommendation rather than just a fact. The real
orphan failure has an empty note cell, is arguably the most consequential line
on its page, and the page was signed as a whole immediately afterwards -- so
the QC flag has to fire at sign/freeze time alongside the blank-result gate,
not in the turnover package where the signature has already happened. ROADMAP
M5.1 records that.

Nothing structural changes. The ten observations still hold, R2/R6/R7/R8 stand
as folded in, R3-R5 remain open. Method note kept in the review notes for
whoever reads a form-heavy source next: render the pages; a text layer is a
convenience, not the document.

Author: mds08011

docs/ROADMAP.md

@@ -472,9 +472,18 @@ gate moves to where ambiguity becomes permanent (ADR 0018 §4).
   Capture itself is never blocked (hard constraint #1).
 - **Turnover.** A **completeness assertion**: the freeze fails while any
   included check item is blank, so a frozen package can never contain an
-  ambiguous blank. The strongest single evidence improvement in the track. QC
-  also flags a **failed line carrying no linked punch item** — the source has
-  exactly one such orphan failure, appearing in no action-item list anywhere.
+  ambiguous blank. The strongest single evidence improvement in the track.
+- **Orphan failures are the same class of defect, and the flag fires at
+  sign/freeze — not at turnover.** A **failed line carrying no linked punch
+  item** is a line whose consequence has nowhere to go, exactly like a blank.
+  The source has one: the line confirming automatic recovery from simulated
+  power-loss events, marked N, **note cell empty**, appearing in no action-item
+  list anywhere in the package — and the page was **signed as a whole**
+  immediately afterwards. A turnover-time check would have caught it far too
+  late; the signature had already happened. So it belongs beside §4's
+  blank-result gate at sign time. Whether it hard-blocks (as blanks do) or warns
+  loudly is the open part — §4's gate is the precedent, and the field evidence
+  says a warning would have been clicked through.
 - **Three questions this item must answer before its task is written**
   (source-reading findings, [ADR 0018 review notes](adr/0018-review-notes.md)
   R3–R5, open questions 6–8):