fix(data): correct version-upgrader self-service guidance + reject corrupt stamps

[krisnye]

Small follow-up to the version-upgrader work (#184). Four issues surfaced in a post-merge re-review — three documentation/guidance fixes and one correctness hardening.

Fixes

1. Drift recipe misdirected toward the removed version resource (assert-versions-match-schema.ts). The self-service recipe printed on schema drift still said "then set the version resource default to N" — but feat(data): ordered version-upgrade system for persisted schemas #184 removed that resource; db.version now derives from entries.length - 1. An agent/human obeying it would re-add a databaseVersion resource, which then registers as an untracked component and re-drifts the guard with no exclusion param left to escape. Dropped the clause (the doc rule already said "there is no resource default to update").

2. data-lit-todo versioning template carried the same dead reference — it's the file every versioned app copies from, so the stale line propagates.

3. readSchemaVersions accepted a corrupt null/missing stamp as version 0 (create-database.ts). Number(value ?? 0) — and Number(null) === 0 — meant a present-but-null (or absent) quadrant stamp read as version 0 and silently triggered a full replay from 0 instead of rejecting. Now only a wholly absent schemaVersions block is legacy (0/0); once the block is present, each quadrant stamp must be a finite number or the load rejects non-destructively (live db untouched). Matches the existing non-numeric→reject contract. +2 tests (null stamp, missing quadrant stamp).

4. version-entry handler doc overstated its guarantee. It claimed the handler "sees a known shape for every component" — false under a scoped/split load, where create-version-upgrader stages only the handler's own quadrant (pickQuadrant). Documented the single-quadrant contract: a handler must read/write only its own quadrant; reading another returns stale/default data under a scoped load and decides the migration wrong. (A runtime read-guard would need store instrumentation on the live path, which we avoid — deferred as a separate option if split-persistence adoption grows.)

Verification

• pnpm -r run typecheck, pnpm run lint, @adobe/data build — all clean.
• Versioning suite green (94 tests), including the two new corrupt-stamp cases.

🤖 Generated with Claude Code