fix(crdt): redeliver rows a late-stamped push changes, and pull counters per node

[juicycleff]

This one stacks on #35 (fix/crdt-sync-defects), because both change the push path in crdt/server.go. Merge #35 first, then retarget this to main.

It fixes edits that never reach other devices. Say device A goes offline, bumps a counter, and comes back a day later. The server merges A's change fine, but B already pulled past A's old timestamp, so B never sees the new total. In the Dart convergence test B sat at 7 when it should have been 10. With this branch it reaches 10.

What was wrong

There were two bugs behind it.

The pull cursor pages by each shadow row's stored HLC. When a push merges a late-stamped change into a row, the state changes but the row kept its older HLC. Any peer whose cursor was already past that HLC skipped the row for good. This hits counters, sets, lists, text and documents. LWW mostly escaped, because an older write loses anyway.

Counters had a second problem. A counter row merged from another node was pulled back as that row's own node delta only, so a peer pulling from scratch still undercounted.

What changed

A shadow row now keeps its own clock (in crdt_state, which is what pulls deliver) apart from its cursor position (hlc_ts, hlc_counter). When a push changes a row's stored state, the server restamps the cursor with a fresh server-clock HLC, so every earlier cursor sorts before it and the row gets redelivered. The state's own clock doesn't move, so LWW and tombstone results come out the same on every replica.

A push that changes nothing writes nothing, so nothing is restamped. That covers an LWW loser, an older tombstone and a retried push. An LWW winner is restamped but keeps its own clock and node. Counters are pulled as one record per node, so a peer always gets every node's delta. The wire shape is the same, there are just more records for a busy counter.

Running it

go test -race ./crdt/...

There are nine new tests covering set, list, text, document and counter late stamps, an LWW loser and winner, tombstones, cursor monotonicity, the multi-table window from #35, and the stream. The ones that check the fix fail on the base commit. The Dart client in #34 also passes its full conformance suite against this branch, with the skipped late-stamp test turned on.

Things you should know before merging

• Pushes now take one lock per plugin, plus one extra indexed query per changed row. That caps push throughput, and a plugin hook that calls back into HandlePush would deadlock.
• If several instances share one database, a peer can still miss a row when two instances commit out of order. The window is much smaller than before, but closing it fully needs a database-side sequence.
• The Go peer Syncer and local AfterMutation writes still put the state's clock in the cursor columns. A node that runs a Syncer and also serves clients still has the bug on that path. That's a follow-up.
• Rows merged from late stamps before this ships aren't restamped. Peers get them on the next change to the row, or on a full resync.