Per-scope takeover for live nodes whose server value is still streaming (D8 follow-up to #3817) #3819

Problem

D8 (Stage 8, documentation/plans/stage8-connection-transport.md) makes a live node take over, meaning it re-runs its compute and opens the live connection, when its own hydration scope releases. For a node in the shell, that's when the root pass ends. Its stated goal is "No live node waits on another boundary."

#3817 (fixing #3764) added an exception. A live node whose serialized value is still in flight when it arms takes over at hydration end, after the last streamed boundary on the page has hydrated. Nodes with a loadingValue are exempt. The fallback is needed because the boundary that reads the node claims its server DOM against that value. If the node takes over before the value arrives, it has nothing to replay, and the boundary can't claim: you get a duplicate or unclaimed DOM. That happens with the real live() transport too, since its local first yield on takeover is the value the node adopted, and there isn't one yet.

The cost is latency. A live source created in the shell and read only inside a streamed <Loading> now goes live only when the slowest boundary on the page hydrates. In practice that covers nearly every real source in that shape: a shell memo read only inside <Loading> misses the shell flush unless its first value is ready synchronously. In a server probe, three microtasks were already too late.

Motivating case: the room demo

examples/room's header creates presence in the shell and reads it inside <Loading>. The Archive panel lower on the page takes 4 s. With the fallback, the tab joins the room, and other tabs see it join, only after the Archive lands. The demo exists to show the opposite. Its comments now note the exception.

Why the client alone can't do better

The ideal gate is "take over once every boundary that reads this node has hydrated." The client can't know that set. When the root pass ends, a pending boundary is an owner with no children, because its children are created when its fragment arrives. So until each boundary hydrates, nothing on the client can tell whether it will read the node. Walking the node's current observers only sees readers that already exist.

Candidate directions

1. Server-recorded readers.
   • During SSR, record for each serialized shell node the ids of the streamed boundaries that read it, and ship that list with the payload, either beside the value or in the readers' fragment data.
   • On the client, the node's gate counts down as those boundaries release, falling back to hydration end, and nodes read by no boundary keep root scope.
   • This preserves D8 exactly, at the cost of a read hook on the server memo path, payload bytes, and an estimated 120–250 B minified on the client.
2. Take over once the node's own value has arrived, replaying it.
   • Keep the per-scope gate, but for an in-flight node also wait until its serialized value lands. Then stamp the takeover with that value (solid.LiveResumeFrom), so the transport yields it locally before connecting and the reading boundary still claims against the server value.
   • A prototype made during fix(solid): a live node with a streaming server value takes over at hydration end (#3764) #3817's review found:
     • Flipping the gate on the settle alone loses a race: the takeover supersedes the adoption before the value commits. Resuming from the landed value fixes that deterministically.
     • It fixes [2.0 rc.10–rc.13] A live source created outside a streamed <Loading> takes over when the shell finishes hydrating, before the boundary's fragment hydrates: unclaimed server DOM + duplicate render, or undefined rows in a keyed <For> #3764's case A and keeps the fast-boundary and loadingValue cases prompt.
     • It relies on the source honoring the replay protocol, so hand-rolled branded iterables still fail.
     • Residual failure: when a value lands before a slower boundary that reads it, and the server's data genuinely changes in that gap, the boundary claims against changed data.
     • As written it added about +183 B minified to hydrating bundles (three scenarios over cap). A tighter version is likely near 100 B.

Either direction should keep #3817's tests (both case-A timings in packages/web/test/hydration/live-shell-source-3764.spec.tsx) and the loadingValue test passing. It should also add a test for a node read by a fast boundary while an unrelated slow boundary is pending, which should connect before the slow boundary lands.