CSS & JS Runtime Audit — Tiered Stylesheets, Dead CSS, a Leaner Entry — 2026-10-09¶
The ask: speed the site up further with a full CSS and animated-JS audit — remove or consolidate classes that are unused or could be combined, enforce a more consistent scheme, and replace or update JS UI libraries where something lighter or faster exists (or another language would do better).
Companions:
performance-audit-2026-10-08.md(frame budget at rest; its Not done list named the render-blockingglobals.cssas “a load-time project of its own” — this is that project) andui-perf-audit-2026-10-09.md(page-switch lag; its Part 2 measured the same sheet as the bandwidth bottleneck on 3G).Finding IDs:
CSS-xstylesheets,JS-xthe entry chunk,LIB-xlibraries.
How it was measured¶
Two production builds side by side —
mainat49a89c3in a git worktree, and this branch — booted from.output/server/index.mjsagainst the same freshly migrated local Postgres.Sizes from the built assets (raw, gzip −9, brotli), and from the repo’s own gates:
scripts/ci/bundle-budget.ts,pnpm run check:bundle-budget(the CI gate) andpnpm run check:entry-compositionon a--sourcemapbuild, which attributes every byte of the entry’s static closure to its source file.Correctness: a computed-style diff, not screenshots. Playwright loads each scenario on both servers and compares 39 computed properties (box, spacing, colour, type, radius, shadow, filters, transforms, animation name) for every element in
<body>, matched by DOM path and class list. 27 scenarios: 13 site pages, 10 full-screen games/apps, and four client-side navigation chains throughwindow.__TSR_ROUTER__.navigate()(site→game, game→site, site→game→site, game→site→game), with a marker onwindowasserting the document survived each hop.
CSS-1 — The site stopped downloading every game’s utilities¶
Before: one entry sheet, globals.css, render-blocking on every page:
488.7 KB raw / 68.3 KB gzip / 51.9 KB brotli. Tailwind’s scanner cannot know
which page a class name ends up on, so it carried every utility any of the ~60
full-screen games and apps use — 93,865 scanned candidates.
An import-graph walk from the site shell (__root, _site, _site/**) shows
1,674 source files that no site page can ever render: 93 components/ and
lib/ directories (every game and app) plus their route files.
After: two entry sheets over the same globals.css body.
Sheet |
Who loads it |
Raw |
gzip |
brotli |
|---|---|---|---|---|
|
every page |
488.7 KB |
68.3 KB |
51.9 KB |
|
|
313.0 KB |
48.5 KB |
38.1 KB |
|
full-screen games and apps |
484.4 KB |
67.6 KB |
51.3 KB |
A site page’s render-blocking CSS is −29% gzip / −36% raw — less to download and ~175 KB less to parse before first paint. Games pay what they paid before.
Why the app sheet is a superset, not “just the game utilities”¶
React 19 never removes a precedence stylesheet once it has inserted one, and
TanStack’s HeadContent renders every head stylesheet that way. So after a
visit to a game its sheet stays in the document. Two partial sheets would then
interleave in an order Tailwind never sorted — a game sheet’s .p-4 landing
after the site sheet’s .px-2 — and silently change what a site page looks
like. A superset that lingers only repeats rules the site sheet already has, in
the same relative order (verified: all 12,195 site-sheet lines appear in the app
sheet in order), so no cascade outcome can change.
The other half of the guarantee is in lib/style-tier.ts: once a document holds
the superset it keeps using it, so the site sheet is never inserted after it
(the game→site→game hazard).
Pieces¶
app/site-tier.css/app/app-tier.css— the two entries.globals.csslost its first line (@import 'tailwindcss') and is otherwise the same body; every test that reads it as text is unaffected.lib/style-tier.ts—styleTierFor()picks per route:_site/**and the root-only 404 take the site sheet; top-level routes take the superset unless listed inSITE_SHEET_ROUTES(/login, articles, embeds…). Forgetting to list a new page costs it bytes, never styles.installStyleTierPreload()warms the app sheet from a site page on three signals the site already acts on: the router preloading an app-tier route (a hovered<Link>, or an on-screen link on a fast connection — the existingviewport-prefetchpolicy, Data Saver included); a client navigation into the app tier starting (the sheet starts with the chunk, not after it); and hover/focus/touch on a plain<a href>into the app tier — see the cost note below for why that one matters most.scripts/fix-ssr-css-hash.mjshandles both entry names.
Gate: lib/__tests__/style-tiers.test.ts¶
The site sheet’s skip list is a manifest, and the test keeps it honest in both directions:
correctness — walks the import graph from everything that takes the site sheet and fails if any import lands in a skipped path (a site page missing its utilities);
leanness — fails if a
components/orlib/directory is imported only by full-screen routes yet still scanned by the site sheet, printing the line to add. It fired during this audit: once/speedrunsstopped importinglib/game/replay(JS-1),lib/daily-puzzlesandlib/lights-outbecame game-only and it asked for them.
Mutation-checked: excluding components/feed makes it fail with the offending
files listed.
Results¶
SSR links the right sheet for every route type, including the 404.
Computed-style diff: 0 CSS-caused differences — 27 scenarios on the default theme at 1440px, and 20 each under Midnight (
.style-graphite), High contrast and a 390px touch phone: ~16,000 element comparisons. The only deltas were sub-pixelwidths onfont-mono/font-seriftext with identicalfont-familyand identical loaded faces, and they swap sides between runs (67.20 vs 67.44 px on one element, either server) — text-rendering timing of thefont-display: optionalfaces, not either build’s CSS.No unstyled frame on a client switch. At 3G (1.6 Mbit/s, 300 ms RTT),
/games→/daily: the app sheet request starts 32 ms after navigation begins, and aMutationObserversaw no/dailyDOM before the sheet had loaded.The cost, stated plainly. Before the split, every page shared one sheet, so a game opened from the site found its CSS already cached. Now the first game of a session needs the app sheet (~68 KB gzip; cached afterwards). Most of those trips are full document loads — the tiles on
/gamesand/appsare plain anchors — so the sheet is fetched on hover/focus/touch of the tile, or earlier when the viewport prefetcher warms an on-screen game link on a fast connection. A cold tap with nothing warmed (Data Saver, or a slow link where the site deliberately doesn’t speculate) downloads it with the game. On the throttled 3G profile a client-side switch with nothing warmed measured roughly +0.1–0.4 s overmain. In exchange every site page — the majority of page views, and every visitor’s first — paints with 20 KB gzip less on the critical path and ~175 KB less CSS to parse.
CSS-2 — 4,000 lines of CSS nothing renders¶
Every class selector in the 55 hand-written sheets was checked against every
source file (TS/TSX/JS/HTML/JSON/MD under app, components, lib, hooks,
stores, server, public, data), with dynamically built names
(`ds-ms__cell--n${n}`, `--rmhfash-swatch-${id}`) and
library-generated DOM (MapLibre’s .maplibregl-canvas) checked separately.
Removed |
Count |
|---|---|
Rules or selector-list entries whose class nothing renders |
537 |
|
32 |
Theme tokens ( |
23 |
Lines |
~4,070 across 23 sheets |
The biggest were leftovers of features that were rebuilt: the old Dunesday
landing page (~1,200 lines, before Dunesday 7’s desktop), the old Lockdown
landing (~790 — lockdown.css is now just the legal-page layout), RMHLadder’s
retired shell and drawer (~700), RMH Vibe’s old gallery/account menu (~320), and
the pre-radial top nav and bottom dock in the shell-wide feed.css (~300; the
entry index-*.css went 12.8 → 8.7 KB).
The removal is selector-aware, which mattered: a dead class inside
:is(.glass-fill, …, .glass-chrome--aside) drops out of the list rather than
killing the rule, and :not(.dead) is treated as always-true. Kept on purpose,
though currently unused, because the design language prescribes them:
.glass-chrome--aside (the escape hatch for position: fixed inside blurred
chrome) and .media-scrim / .media-scrim-full. The display scale lost its two
unused hero steps (--site-display-1/2); the comment says how to restore them.
design-consistency.test.ts’s CSS_DEBT ratchet lowered five entries to match
(the prune removed hardcoded colours and radii with the rules).
JS-1 — What every page’s entry carried that it never runs¶
A --sourcemap build attributed the entry’s static closure by source file. Every
route definition is in it (routeTree.gen.ts imports all of them; Start’s
splitter lifts out only components), so anything a definition touches at top
level runs on every page. Four such leaks, all fixed:
Leak |
Size (min) |
Fix |
|---|---|---|
zod, via three route |
~70 KB |
Titles and version tags → zod-free |
RMHbox minigame registry, via |
~10 KB |
The description no longer carries the count. |
RMHbox history registrations, a bare |
~8 KB |
|
The whole |
~9 KB |
Un-exported. |
zod was the one dated exception in check-entry-composition.ts’s
KNOWN_VIOLATIONS ratchet since 2026-08-05; it is gone from the entry, so the
baseline is deleted and zod in the entry is a plain failure again.
route-css-imports.test.ts gained two rules for the other two shapes (no bare
side-effect import of a module in a route file; a page route exports only its
Route), mutation-checked.
Critical path ( |
|
after JS-1 |
after JS-2 |
|---|---|---|---|
Eager JS, brotli ( |
303.1 KB (over its 300 band) |
278.3 KB |
242.7 KB |
Entry chunk, raw (CI gate, budget 525.3 KB) |
628.6 KB ✗ |
594.1 KB ✗ |
460.2 KB ✓ |
Critical path, raw (CI gate) |
1,202.3 KB |
1,090.5 KB |
956.6 KB (−20%) |
Critical path, brotli (CI gate) |
336.6 KB |
309.9 KB |
274.4 KB (−18%) |
JS-2 — No translation catalog on the critical path, English included¶
After JS-1 the CI gate’s entry raw row was still red (594 KB against 525 KB —
red on main too, at 628 KB). The largest single item left in the entry was the
English core catalog, ~136 KB minified: 13 namespaces (feed alone is
74 KB of JSON) bundled so the shell and feed could hydrate in English. The
sourcemap charges it to react-i18next/I18nextProvider.js, which is why the
composition table read “react-i18next 138 KB”.
Nothing on screen needs it. Every t() call carries a defaultValue, so
English can render from those on the server and on the client’s first render
alike — hydration matches by construction — and the catalog can arrive at idle,
as the non-core English catalog already did. That is safe only if every default
says exactly what the catalog says, so that the catalog landing changes nothing.
It did not, quite:
Found |
Count |
Fix |
|---|---|---|
Core defaults that drifted from the catalog (casing, |
33 |
Default aligned to the catalog — the text users already saw, so nothing visible changes |
Plural calls passing one default for both forms |
4 |
|
English catalog plurals that were wrong (“Show 3 more reply”, “3 new post”) |
2 |
|
|
2 call sites |
Pass |
Template-literal defaults ( |
17 calls / 19 keys in |
Interpolated defaults ( |
Then resources.en-core.ts is deleted and no catalog is bundled: the server
renders English from defaults (getServerI18n), the client initialises with no
English resources, and backfillEn() loads the whole catalog after load +
idle — for other locales’ fallback and anything read without a default.
Gates. i18n-default-drift.test.ts now covers every namespace (it used to
exempt the bundled ones — the 41 drifts were hiding there) and gained a rule:
every call with a literal key carries a literal default (a template-literal
default is reported as none, with the reason). i18n-code-splitting.test.ts
lost its resources.en-core exception. Both mutation-checked.
Verified at runtime, because 58 core calls build their key from data and no
static check can vouch for those: each page rendered twice from the same build —
once with the English catalog request blocked (what a visitor sees before the
backfill) and once with it loaded — diffing every visible text node plus
placeholder, aria-label, title and alt, with the command palette and the
shortcuts sheet opened on each.
Pass |
Renders |
Strings |
Differences |
|---|---|---|---|
Signed out, 34 pages × 1440 + 390 px |
68 |
~21,900 |
0 |
Signed in, 16 pages (settings, messages, notifications, wallet, profile…) × 2 widths |
32 |
~12,400 |
0 (one render came back signed-out and was re-run clean twice) |
German locale ( |
6 |
~2,200 |
0 — German renders server-side, English fallback unchanged |
No new hydration errors. Five pages (/games, /apps, /store,
/leaderboard, /pricing) log React #418 with or without the catalog — and on
the PR’s previous commit, and on a fresh build of main. It is pre-existing: on
/games the server renders “Loading your standings…” where the client renders
the signed-out state, and the stack points into the subtree
DeferOnNavigate wraps. Out of scope here; reported separately.
The eager-CSS budget was measuring the wrong file¶
platform_eager_css (40 KB brotli) summed only the CSS the manifest attaches to
the entry chunk — the 3 KB index-*.css. The real render-blocking sheet is
linked from __root’s head through a ?url import, so no manifest lists it and
the budget never saw it. It now counts the site-tier sheet. Honest figure:
~40 KB on this branch, where main would read ~55 KB.
LIB — Libraries: what was replaced, and what was not¶
zod — off the critical path (JS-1); still used by server functions and game/app chunks, where it belongs.
uuid,lodash-es,immerlook unused by the app, but are resolved from ournode_modulesat runtime to bundle user-made RMH Vibe pages (lib/rmhvibe/vibe-packages.ts, tierinline). Removing them would break published Vibe pages. Kept.framer-motion (
motion-dom+framer-motion, ~42 KB min in the entry, behindLazyMotion) — 108 files use it. A wholesale swap is a large rewrite for a cost the entry composition does not single out; the 08-12 audit already moved scroll reveals to CSS. TheLiquidTabslayoutIdthumb the previous audit flagged was looked at and left: since NAV-1 its layout read is the commit’s first, so it duplicates nothing, and a CSS-only thumb would either paint no selection until JS measured it (losing the server-rendered state) or need the same forced read for a FLIP.tailwind-merge (26.5 KB) —
cn()’s override semantics are relied on by every primitive that acceptsclassName;clsxalone would let a default win over a caller’s override. Kept.sonner (32.8 KB) — mounted globally;
toast()and<Toaster>live in one module, so lazy-loading the toaster saves nothing. Kept.Another language? Nothing on this path qualifies. CSS is parsed by the browser; the entry’s JS is React, the router and i18n. The one measured WASM candidate remains the Slice It STFT, where
worker_threadsis the first fix (08-12 audit).
Not done (and why)¶
React #418 on five pages, pre-existing on
main(JS-2 above). A session-dependent render mismatch, not a size or i18n issue.App-route-only rules still in
globals.css(.app-viewport,.app-page,html.app-route …, ~2.5 KB raw): some top-level pages that take the site sheet also gethtml.app-route, so moving them needs a per-route audit for ~0.5 KB brotli.