perf(realtime): page boot as the shared runtime — islands carry thin hooks (~−50 kB on /posts/:id)

[sebyx07]

Follow-up to #504 (plan 101), written from the plan-101 bytes audit. Every island bundles its own copy of the page's realtime runtime, because islands build with `splitting: false`.

Pair on one page	Shared bytes
like ∩ badge	26,444 (98% of the badge)
like ∩ boot	26,833
feed ∩ boot	26,833

Cuts, ranked

1. Make the page boot the shared runtime. Move the store, socket client, host, channel book and outbox into the boot, and have islands reach them through the existing `Symbol.for('ultimate.realtime')` global with thin hooks. This is decision 12's "next lever" without a splitting build.
   • `/posts/:id`: about 145 kB → about 95 kB.
   • `/feed`: about 133 kB → about 98 kB.
2. Browser error tables: only codes a browser can raise. About 5–7 kB per live island. Today `core-error-codes`, realtime `errors.ts`, `page-errors` and schema `errors` all reach island chunks.
3. In-page socket engine only on fallback. About 5.6 kB per live island. `socket-engine` + `socket-routes` run only without a SharedWorker. Load them as a framework script by URL (the worker bundle already contains them).
4. Seed the store from the server-rendered row. This needs a record-adopt seam and drops one GET plus about 2.2 kB from the like island.

Owners: realtime (`page-socket`, `use-*`, `boot`) and cli (`island-realtime`, `worker-bundle`). Internal only, so it can ship in a minor.

[developerz-ai]

Great follow-up to the plan-101 audit — this reads as a clear enhancement rather than a defect: the realtime runtime already lives on the page, it just rides along inside every island bundle. The intersection table makes the prize concrete (~50 kB off /posts/:id, ~35 kB off /feed), and the ranked cuts give the work a natural order — boot-hosted runtime first, then the browser-only error tables, the fallback-only socket engine, and store seeding last since it needs the record-adopt seam.

Labeling this enhancement and putting together a scoped plan next.

🤖 Posted by developerz.ai — the maintainer agent, not a human.

[developerz-ai]

Scoped this so it's reviewable before anyone picks it up — the cut that carries the win is #1 (boot-hosted runtime); the rest are separable follow-on levers.

Problem — every island bundles its own copy of the page's realtime runtime (~26.8 kB of shared bytes per island, ~98% of the badge island), because islands build with splitting: false. The page boot should host the runtime; islands should reach it through the existing Symbol.for('ultimate.realtime') global via thin hooks.

Acceptance criteria

• Islands no longer bundle the store, socket client, host, channel book or outbox — they reach them via the Symbol.for('ultimate.realtime') global through thin hooks.
• /posts/:id page JS drops from ~145 kB to ~95 kB (or a comparable measured reduction in the bytes audit); /feed from ~133 kB to ~98 kB.
• Realtime behavior (likes, badges, feed updates) works with no functional regression in tests.
• Build and existing test suites pass in CI; bytes audit updated to the new sizes.

Areas — packages/realtime/src (boot, page-socket, use-* hooks, store, socket-engine) and packages/cli/src (island-realtime, worker-bundle), plus core-error-codes for the error-table trim.

What a fix should not do

• No splitting: true island build as a quick win — the boot-runtime seam is the point.
• No new ad-hoc window global — the Symbol.for('ultimate.realtime') seam already exists.
• No deleting or skipping bytes-audit checks to make size gates pass.
• No shipping both copies behind a feature flag.
• No weakening the Symbol seam's typing to any.

Cuts 2–4 (browser-only error tables ~5–7 kB, fallback-only socket engine ~5.6 kB, store seeding via the record-adopt seam) fit the same shape and can land on their own once the boot seam is in.

🤖 Posted by developerz.ai — the maintainer agent, not a human.

[developerz-ai]

Want to submit a fix? I'll scope it and flag it for the maintainer to pick up — the plan above is ready to work from, and cut #1 alone is a well-bounded first slice if you'd rather land it in pieces.

🤖 Posted by developerz.ai — the maintainer agent, not a human.

[sebyx07]

Plan 101 slice 08 row s: contact-sales ships ~6.3 kB of error tables in a non-live island. This issue's browser-only error tables should cover action-only islands too, not just live ones.

[developerz-ai]

Good catch — and it's the structurally expected shape: nothing in the island build gates error tables on live-ness. Islands bundle with splitting: false, so whatever the island's import graph touches gets inlined, and a contact-sales form island that reaches the action client pulls the same core-error-codes / realtime errors.ts / page-errors / schema errors tables a live island does. The "~5–7 kB per live island" figure was just where the audit happened to measure.

Folding this into the working scope for cut #2 above:

• The trim — keep only codes a browser can actually raise — applies to every island build, live and action-only alike. The ~6.3 kB on contact-sales is the non-live data point.
• Acceptance wording shifts from "per live island" to "no island of any kind ships browser-unreachable error tables."

One addition worth putting on the list: the bytes audit today measures live islands (/posts/:id, /feed), which is exactly how a non-live island's 6.3 kB stayed invisible. Adding one action-only island row to the audit keeps this class measured after the trim lands.

The scoped plan above stands with that amendment, and the offer above still stands either way.

🤖 Posted by developerz.ai — the maintainer agent, not a human.