Prove portable SQLite claim and BLOB behavior

[kcsfelty]

Summary

• add one shared low-level SQLite capability suite, executed through four persisted Node SQLite connections and a real Miniflare/workerd D1 binding
• prove the observable conditional semantics of UPDATE ... RETURNING: concurrent submissions produce exactly one winning claimant, and a revision-only stale probe produces zero rows/changes
• retain the existing separate child-process test as the evidence for native physical lock contention and bounded busy timeout; the shared suite does not claim to force lock overlap or scheduling order
• prove scalar and BLOB bind/read round-trips in both runtimes
• add the narrowly required readSqliteBytes() helper because Node returns Uint8Array while workerd D1 returns a JSON-compatible byte array; reject sparse and malformed byte arrays
• document the Diamond handoff boundary: no search, ranking, auth, cursor, queue, lease, or vector semantics

Evidence

• npm run check (165 tests, coverage 96.24% statements / 91.94% branches / 97.74% functions / 96.24% lines; build, bundled workerd, packed consumer, license and readiness gates pass)
• npm run conformance (490/490 applicable W3C SPARQL 1.1 cases)
• shared capability suite passes in persisted Node WAL and real workerd D1
• existing bounded child-process contention test remains green

Strongest counterargument

The new helper expands the production API for what could remain an application-local conversion. The runtime evidence shows the row shapes genuinely differ, however, and leaving normalization to each consumer would undermine the shared adapter contract. The helper is deliberately only a dense, validated, copying byte conversion; it contains no domain behavior.

Production API impact

One additive root export: readSqliteBytes(value): Uint8Array. No existing behavior or type changed.

CI notes

The first W3C run passed its 490/490 summary and then failed the second EARL pass on a transient upstream manifest fetch; a fresh run is pending. The audit gate independently began failing on a newly published high-severity sharp/libvips advisory in dev-only Miniflare/Wrangler. Latest upstream Miniflare still pins vulnerable sharp@0.34.5; this PR does not force an unsupported major override or conceal that advisory.