Scope graph rebuilds per store, on master's page enumeration

[alistair3149]

Supersedes #1233, rebased onto master and reworked onto the page enumeration #1226 introduced. #1233 stays open as the record of the original design and its review; close it once this is accepted.

CI is green: 2,670 tests across MW 1.43–master, PHP 8.3–8.5.

Why this exists rather than a rebase of #1233

#1233 predicted this: "whichever merges second needs rework, not conflict resolution." #1226 merged first, and it did not just rename things — it changed what a rebuild reconciles:

	master (post-#1226)	#1233
Page enumeration	PageIdsLookup::getPageIds( $afterPageId ): iterable	SubjectPageIdsLookup::getSubjectPageIdsAfter( $after, $limit ): array
Deletion enumeration	DeletedPageIdsLookup::getDeletedPageIds(): iterable	getDeletedSubjectPageIdsAfter( $after, $limit ): array
Rebuilder	PageRebuilder	SubjectPageRebuilder
Scope	every page	only pages carrying a Subject

Resolved mechanically, the tree would have held two enumerations of the same wiki and a rebuild that reconciled a subset of what every other path projects — so a rebuilt graph would no longer match a saved one.

What changed

The rebuild walks master's lookups; the four Subject* duplicates are gone. Both of master's contracts gain what a resumable rebuild needs and an unbounded walk does not: a bounded read returning one batch, so a run can record how far it got and continue there rather than part-way through a generator, plus a count to report progress against.

This PR also carries the 13 review-fix commits that were #1250, folded in before the rework so the same conflicts were resolved once rather than twice.

Decisions worth a look

1. The deleted-page cursor. master's getDeletedPageIds() pages by ar_id because archive has no index on ar_page_id, and sorting it per batch is exactly what that avoids. A run resuming from a page-id cursor cannot start from an archive id, so the new bounded read orders by page id and pays that sort — the cost of being resumable, and the same cost Scope graph rebuilds per store, batched and resumable, with run records #1233's own deleted-subject lookup already paid. master's unbounded reader keeps its faster walk. The alternative is threading a second cursor concept through RebuildProgress and the run record.
2. PageUndeleteComplete had two handlers — master added one for reprojection, Address the review findings on the scoped graph rebuild #1250 added one for Mapping changes. Merged into one that does both.
3. The skip reason was wrong after the scope change. The executor logged "it carries no Subject to project" for every non-Refreshed outcome, which is now false: such a page is projected like any other. It reports PageRefreshOutcome::skipReason() instead.

What the tests needed

The scope change reached further into the suite than the production code did, and all of it was rebase damage rather than defects in the original work:

• Nine assertions were arithmetic. A test's Subjects need a Schema page, and that page is now projected too — created first, so it sorts ahead of the pages a test names and fills the first batch alongside them. Counts gain FIXTURE_PAGES; lists are bounded by the first page the test made, via a helper on the shared base class.
• Three store-death tests were no longer testing anything. The rewind fires only on a wholly refused full batch, and the fixture's page shares batch one and always projects successfully. Left alone, those tests would have gone green while exercising nothing; the refused pages move up one so the batch they target is full.
• Two tests changed meaning. A page carrying no Subject is projected rather than skipped, so testAPageWithNoSubjectLeftToProjectIsSkippedAndWalkedPast lost its premise and now asserts what happens instead. And saving a Mapping page projects it into stores the test points at unreachable endpoints, so an assertion that the error log was empty now filters to errors about the rebuild.
• Three helper collisions were fatal at load. This branch moved helpers onto NeoWikiIntegrationTestCase that master's tests also declare, and PHP will not let a subclass narrow an inherited method — the suite died before a single test ran. One was a true duplicate; two were different things sharing a name.
• One test file lost its helper in a rename conflict. master renamed DatabaseDeletedSubjectPageIdsLookupTest, so git merged both versions and the resolution kept this branch's tests without master's newLookup(). Its remaining Subject-scoped assertion went too: the lookup reports every page the wiki lost, which is what the test's own name says it checks.

Still open for review

Two semantics questions the suite passing does not settle:

1. wasOfferedToTheStore under every-page scope. It exists so a page the walk found but could not offer does not count toward "every offered page failed" in the store-death probe. With no-Subject pages now offered, the unoffered set shrinks to unreadable pages, so MIN_BATCH_SIZE_FOR_STORE_DEATH and the probe's guard deserve a fresh read.
2. Skipped pages are invisible in the sync verdict. A page whose slot content will not parse (SkippedUnreadableSubjects — the exact case the hook path warns about, telling the operator to run this script) is skipped at info level, not counted, and the run reports "in sync" with exit 0 while the store provably lacks that page. Counting unreadable skips as failed, or carrying a skip count on the run record, are the options; both change semantics, so this is a design call rather than a patch.

Verification

make cs clean; CI green across MW 1.43–master on PHP 8.3–8.5 (2,670 tests, 7,197 assertions).

_{AI-authored — Claude Code, Opus 5 (1M context); rebase, rework and review fixes in-session; not yet human-reviewed.}