docs: state the per-page temporal replay boundary in its true form (#142)

[Xof]

Closes #142 (I153). Stacked on #159. Documentation only — 731 tests still pass.

The design spec's §9 documented "no rollback/replay resistance" for one case: an attacker who substitutes a wholly older, validly-signed database image. The actual weakness is strictly stronger, and the gap between the two mattered.

Page AAD is page_id and nothing else, so a sealed page authenticates where it belongs but not when. An attacker holding an older copy of the file can splice individual stale pages into the current one — each verifies under the current DEK, at the correct page id. What comes out is a mixed state that never existed at any commit: a stale freemap or interior tree page beside current siblings, which the engine may walk as though it were consistent.

Neither of the spec's Provided claims covered this, and both read as though they did:

Anti-relocation is about moving a page to the wrong place. Replay is about moving it to the wrong time. Both listed guarantees are spatial, not temporal.

§9 now says that in as many words, since the distinction is the whole finding.

Stated in three places, for three readers

Where	Reader
Design spec §9	the normative boundary
THEORY.md	someone who will never open a spec
PageCipher::seal doc comment	the maintainer who might "simplify" the AAD, or read anti-relocation as more than it is

What bounds the exposure

• rekey (feat: bulk DEK rotation — re-encrypt a whole database under a fresh data key (#140) #159, ADR 0018) does not make a splice detectable, but it invalidates every page image sealed under the old DEK — so pages captured before a rotation cannot be spliced in after one. An operator who suspects file access has a way to draw a line.
• The I148 positional type checks (fix: type-check membership and handle-table pages against their tree position (#141) #158) are a partial and accidental mitigation: a splice is caught whenever the stale page's type differs from the one expected at that position. It is silent only for a same-type splice — realistic for hot structural nodes that are repeatedly COW'd in place, but far from universal.

Three routes to actually closing it are listed with honest costs (external trust anchor / commit epoch in AAD / hash-chained page tags); none is implemented, and the issue stays the tracker for that.

Two corrections from review

• I had a causal claim backwards. The first draft justified rejecting a commit-epoch AAD as "the write amplification shadow paging exists to avoid". Shadow paging accepts write amplification as the price of crash-safety-by-inspection — ADR 0001 lists it as a negative consequence and THEORY.md says the cost is taken with eyes open. The real objection is amplification well beyond what that trade already buys, which is a different sentence. Fixed in both places.
• "Corrupts structure silently" was stated flatly in two of the three files and hedged in the third. Hedged consistently now, with the type checks named as the reason it is only sometimes silent.

ADR 0015

It stated both the old boundary and "bulk DEK rotation deferred". It is Accepted, so it gets an ## Addendum (2026-08-04) pointing at ADR 0018 and the corrected boundary — not an in-place edit.