Stop the whole dashboard waiting on one query before anything can fetch

[ThinkingSpade]

First fix from a production performance measurement. It removes one of three
waves, not the per-call cost.

What I measured on prod (2026-07-31, signed in)

what	time
/favicon.ico — static asset, Worker never runs	0.07s
/api/does-not-exist — 404, no handler, no DB, no auth	4.16s
/api/auth/get-session	4.04s
Keywords page, last server function finishes	15.8s

A 404 costs the same as a real request, so ~4s is fixed per-invocation overhead
before any of our code runs. Ten back-to-back calls: 3.4 / 4.0 / 5.7 / 5.1 / 3.9
/ 4.2 / 4.6 / 4.6 / 3.4 / 3.8s — not one fast, so there's no isolate reuse to
wait for.

The page fires 14 server functions in three sequential waves (~1.7s, ~5.7s,
~11.5s). The waves are what multiply the 4s into 16s.

The change

The dashboard put every card behind projectsQuery.isPending. Since that query
costs ~4s, no card could begin fetching until 4s in — turning independent
queries into a second wave.

projectId comes from the route params, not from that query. Seven of the nine
cards are keyed only on it, and each already renders its own skeleton and
degrades independently — which is why removing the page-level gate leaves no
holes.

The two that genuinely need domain (AnalyzeProjectCard, BacklinksCard)
still wait, rendered in place behind a skeleton so the layout doesn't reflow.

DashboardLoadingState goes with it — the header now has its own skeleton and
every card brings its own, so it had no caller left. knip caught that.

What this does not do

It does not fix the 4s. And the obvious fix for that was already tested and
falsified on 2026-07-29: cutting the startup chunk 30% (5,933 → 4,166 KB)
moved it not at all. It isn't parse time — the remaining suspect is top-level
module evaluation (agents, think, ai, better-auth, autumn all
initialise at import). Please don't attack it with more bundle shrinking.

On Supabase, since it prompted this

It would fix none of this. The session table has 2 rows with a unique
index on token; a 404 touches no database and costs the same 4s. An
independent Codex pass agreed and added a detail I'd missed: the existing
Postgres path uses Hyperdrive with max: 1, so Promise.all sections would
queue on one connection — potentially making parallel work slower, plus
regional latency.

Verification

pnpm ci:check clean, 2,129 tests passing.

The wall-clock win is not yet proven — it needs a production re-measure after
deploy, using exactly the numbers above. That's the next step, not an assumption.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Updated (UTC)
✅ Deployment successful! View logs	flyrocketseo	c72bdc0	Jul 31 2026, 07:16 AM