Reach a page area a break names, whose <occur max> is spent

[tannevaled]

Six forms of the corpus lay their body on sheets pointing the wrong way, and it is one rule.

ca-cra t2042, t2121 and t2125, each in two years, write a portrait Pg1 with <occur max="1"/> and a landscape page area beside it, and every body subform carries <breakBefore targetType="pageArea" target="Pg1" startNew="1"/>. This package refused the target once Pg1's max was spent and let the sequence answer instead — the landscape page area, 612 points tall where the form wants 792 — so 403 leaves fell off the bottom.

pdfium reads a page area's <occur max> in exactly one place: the branch of GetNextAvailPageArea that REPEATS the page area in hand (cxfa_viewlayoutprocessor.cpp:1552-1578). Every other route runs through FindPageAreaFromPageSet_Ordered, which scans the siblings after the page area in hand FOR THE TARGET and, finding none, starts the page set again from its first child and takes the target there (:1582-1607). That function never reads a page area's occur; the only max it reads is the page SET's (:1329-1341). An instrumented pdfium says the same at runtime.

So a page area's max bounds how often the SEQUENCE runs into it, not how often a break may ask for it by name. This package already held the other half of that rule — starting a page set again forgets the counts of the page areas below it, pager.cleanKids — and pager.reach is that same restart, asked for by a target.

What is not changed is the break that names nothing. Two thousand of the corpus's two and a half thousand breaks write no target, and there pdfium's scan accepts any page area, so the sequence moves on and a max still means something. Getting that wrong first cost 2 367 leaves on four other forms, which is why breakTo now carries whether the template NAMED it.

Measured, 559 forms, hidden excluded from both sides

	before	after
body leaves off the paper	1378	1000
forms with one	37	34
sheets whose SIZE disagrees with pdfium	13	7
...of those, an ORIENTATION difference	6	0
agree on x	173806/177676	unchanged
agree on sheet+y	162200/177676	unchanged
corpus sheets	3088	3088

Six forms move and none is worse: t2125-24e 129 → 0, t2042-24e 94 → 0, t2121-24e 17 → 0, t2125-25e 178 → 66, t2121-25e 96 → 71, t2042-25e 67 → 66. The remainder on the 25e forms is this package's text measurement, not this rule. t2125-25e's page-area sequence is now pdfium's, sheet for sheet.

The judge did not move by one leaf, and that is the finding about the instrument: it compares x, and sheet+y. A leaf at the same place on a sheet of the wrong ORIENTATION counts as full agreement. Nothing in this repository compares a sheet's size.

Nine forms carry a pageArea <occur max> a break names — the six above, ca-cra t1135 in two years and us-uscis i-612 — and 60 carry one at all.

Also

doc.go's corpus census quoted 81 734 fields on 3 090 sheets; both went stale before v0.16.0. It is now what TestPlacementOverTheCorpus reports: 81 750 placed, 151 323 draws, 3 088 sheets, none left over, and the 12 furniture elements that belong to a page area no sheet of their form ever is. Its own commit.

Gate

gofmt clean, go vet clean, -race green, 100% statement coverage with XFACORPUS unset, CGO_ENABLED=0, GOWORK=off, Go 1.27.1 via GOTOOLCHAIN, and 13 cross-compile targets. The full suite is green against the corpus, pdf.js and pdfium references, before and after.

Two tests changed their expectation with the rule and are renamed for it; three are added — the untargeted break that must still move on, a page set whose own <occur> is spent, and a <pageArea> written outside every page set, which a target reaches by id and which has no page set to start again.

Diagnosis for issue #7 is in the issue comment; findings and scripts at /Users/Shared/xfa-offpaper/.

🤖 Generated with Claude Code