0585 — the claim path's snapshot gate proves on PGlite, so the sandbox can tick

Context

The consumer-flow sandbox extends the console [0584] into the whole produce/claim path, and its Tick control depends on machinery built for a condition PGlite cannot create. freshClaimMessagesWithCursor never advances claimed to the raw MAX(id) — an id issued by a still-open producer transaction would be skipped forever. It advances only to a head it can PROVE, by reading a (head, xmax) pair before its transaction’s first write and then requiring pg_snapshot_xmin to have passed that xmax.

That proof assumes many concurrent backends. PGlite is one. If the gate never proved, Tick would claim nothing and every consumer card’s narration would have to change — so this was settled by measurement before any of the flow was drawn.

Decision

Build the sandbox’s tick on the real claim path unchanged. A headless round trip against PGlite in Node — register group, produce, claim, read, commit — proves on the first poll: snapshot pair (head=3, xmax=789), pg_snapshot_xmin also 789, the fresh-pair branch fires, the cursor advances (0, 3], the lease frees under its token and settled_head lands at 3.

It proves structurally, not by luck. The snapshot statement is a pure SELECT, so the transaction holds no xid when it reads xmax; PGlite’s single backend means every producer transaction has already committed by the time the gate CTE runs, so nothing holds xmin down. The settled_head fallback — the branch that carries claims when neither pair proves — is unreachable in the browser.

Consequences

• A produce lands in the very NEXT claim, (1, 2], not a tick later. No card needs a “wait for another tick” caveat.
• Two claims on one group’s cursor come back disjoint, (0, 2] then (2, 3] with first.high == second.low — which is exactly what the shared-group consumer card tells the reader.
• Caught up reads low == high, the same fact the Go code turns into a nil claim, so Tick has a defined no-op state.
• The page can explain the fence but can never show it firing: staging it needs a second backend holding a write open. This is the same one-connection limit the sandbox already discloses for contention.
• The spike’s SQL was pasted from the Go sources, not extracted, and was deleted once it answered the question. The drift-checked extraction and its standing coverage ride with the tick integration step, under [0584]‘s rule.