With serverFunctions.components on, data server functions called during hydration never reach the network (frames intercept waits forever)

Start in packages/web/frames/src/client.ts, where the intercept passed to createServerComponentHandler (around line 1683) calls boundaryMayArrive() and awaitBoundary(id), then read the arrivals cache. Also read subFetch in packages/solid/src/client/hydration.ts and the tracing check from #4013. Reproduce by loading examples/room /live with components: true. Done means data server functions called during hydration reach the network and the summary leaves "summarizing…".

In an app with serverFunctions: { components: true }, a data server function (one that returns data, not a component) called during hydration never sends its request. The frames client's intercept parks the call waiting for a <solid-frame> boundary that a data function never has, and in two ways that wait never ends.

examples/room's /live page shows it on next (4937f0a):

• Never connects: there are no /_server/live/* or /_server/data/summary requests at all, and no console errors.
• UI stuck: the status pills stay at "connecting", presence shows "0 here" and "Not in the room yet", and the summary stays at "summarizing…".
• Writes go one way: a message posted from one tab reaches the server but never shows in another tab.
• Everywhere: it happens in dev over HTTP/1.1, in dev over HTTPS/HTTP/2, and in the production build, while _$HY.done === true.
• Controls: room's / page, a live server component, works on the same server. The same /live source served with components: true removed connects, with both tabs in presence and posts crossing. examples/board doesn't set components: true, which is why it connects on the same build.

This is related to #3819, which uses the same page as its motivating case, but that issue is about the connection opening late. Here it never opens.

Cause

In packages/web/frames/src/client.ts, the intercept passed to createServerComponentHandler (around line 1683) runs for every server-function call, data functions included. While boundaryMayArrive() is true, meaning _$HY.done is unset or a fragment is pending, it treats a function with no boundary element as "a boundary that may still arrive" and returns awaitBoundary(id). That is a promise cached per function id in arrivals, and the call doesn't go to the network. A data function never gets a boundary, so it can only be released by the reveal hook. Two defects keep that release from ever happening.

1. The cached arrival promise can be a MockPromise that never resolves (confirmed)

• The first call happens inside the hydration trace (subFetch in packages/solid/src/client/hydration.ts), which swaps the global Promise for MockPromise.
• So new Promise(r => (resolve = r)) in awaitBoundary creates a MockPromise and resolve stays undefined.
• In the browser, the arrivals entries for presence, transcript, roomCard and archive all had typeof resolve === "undefined", created exactly when Promise !== native.
• The takeover calls after the root pass hit the same cached dead entry and hang.
• If a reveal ever tries to release them, it throws TypeError: arrival.resolve is not a function. That throw was reproduced.

This is the same class of bug as #4013 (a lazy() loader promise made under hydration tracing). That fix added a tracing check, and frames doesn't consult it.

2. Waiters registered after the page's last reveal are never released (confirmed)

The summary source hangs even though its arrival was created with a real Promise:

• Waiters are released only from a reveal notification, and only if hy.done && !hy.fr.pending() holds at that moment.
• Timeline: reveals at about 520 ms and 720 ms, then the last one (the 4 s Archive panel) at about 4015 ms, with done still unset and pending() false.
• _$HY.done = true is set about 5 ms later, from a setTimeout in drainHydrationCallbacks. So the last reveal kept the waiters, and no reveal comes after it.
• Calls made by hydration-end recomputes (summary uses ssrSource: "client") register in that gap and wait forever.
• Firing a synthetic reveal after done (with the dead MockPromise entries dropped first) immediately sent GET /_server/data/summary, and the summary started streaming.

A fresh call after hydration goes straight to the network, because boundaryMayArrive() is false once done is set. Only calls made in the hydration window are lost, and any components: true app that calls data server functions during hydration should hit this.

Possible fixes (untested)

• Don't defer non-component functions in the intercept: a data function can never have a boundary.
• Don't create or cache arrival promises while hydration tracing is on, as #4013's fix does for lazy().
• Release outstanding waiters at hydration end (onHydrationEnd) as well as on reveals.

Regression?

Unknown. The intercept, awaitBoundary and boundaryMayArrive code predates the held-swap policy note (#2964) in the comments, and an older build (around 06c7717dd) behaves the same.