01 — Measurements: harness, numbers, and what not to trust

Everything quoted in this folder came from one of three places. Knowing which matters, because they have very different error bars.

Source

Determinism

Use it for

Build output

Exact

Chunk sizes, critical-path composition, what exists

CDP network trace

±5%

Request counts, encoded bytes on the wire

Browser timings

Noisy

Direction of travel only — see §4

1 — How to reproduce

# 1. Deps and a database (Postgres 16 local, empty).
pnpm install
createdb rmh
# `prisma db push` is blocked for agents by Prisma's own guard. Read-only DDL
# generation + psql apply is the non-destructive equivalent against an EMPTY db:
pnpm exec prisma migrate diff --from-empty --to-schema prisma/schema.prisma \
  --script -o /tmp/schema.sql
psql -h 127.0.0.1 -U postgres -d rmh -f /tmp/schema.sql     # 323 tables, 0 errors

# 2. Production build, then boot it.
pnpm build
PORT=3201 NODE_ENV=production \
  DATABASE_URL=postgresql://postgres:postgres@127.0.0.1:5432/rmh \
  BETTER_AUTH_SECRET=... BETTER_AUTH_URL=http://localhost:3201/ \
  node .output/server/index.mjs

# 3. Measure. Both scripts are throwaway; the source is inlined in §5 below.
node measure2.mjs / /games /apps /library /news

Chromium: use the pre-installed binary, not Playwright’s expected one — the repo pins Playwright 1.62 (wants chromium-1234) and the image ships chromium-1194:

chromium.launch({ executablePath: '/opt/pw-browsers/chromium-1194/chrome-linux/chrome',
                  args: ['--no-sandbox', '--disable-dev-shm-usage'] })

All external origins are aborted at the route level, so every number below is pure client + origin cost with zero third-party network time. A real connection only adds to these.

2 — Build-level: the critical path

The critical path is the transitive static-import closure of the client entry chunk (the one containing rmh-user-theme), walking from "./x.js" and skipping import("./x.js").

Metric

before

after

change

Entry chunk

277.2 KB

276.0 KB

−1.2 KB

Critical path, raw

1323.3 KB

1182.1 KB

−141.2 KB

Critical path, brotli

380.2 KB

338.1 KB

−42.1 KB

Chunks

131

108

−23

Chunks under 6 KB

109 (137.6 KB)

89 (125.9 KB)

−20

Total JS files emitted

1090

1018

−72

Reached in two steps, worth separating because the second was found only after the first had been measured and written up:

Step

raw

brotli

chunks

baseline

1323.3 KB

380.2 KB

131

icons + zod-catalog + socket.io

1260.6 KB

361.4 KB

111

+ framer-motion (9 modules)

1182.1 KB

338.1 KB

108

Largest critical-path members, after:

Chunk

Raw

Brotli

What it is

Needed on every page?

index-*

276.0 KB

66.5 KB

the entry

inherent

vendor-react-*

187.3 KB

50.1 KB

React

yes

c-ui-*

128.3 KB

34.0 KB

English i18n core namespaces

yes

schemas-*

71.0 KB

16.6 KB

zod

no — see §3

localeStore-*

46.3 KB

13.5 KB

locale store

yes

auth-client-*

43.0 KB

13.0 KB

Better Auth client

yes

router-*

41.4 KB

12.2 KB

TanStack Router

yes

createServerFn-*

36.8 KB

10.2 KB

Start server-fn runtime

yes

Providers-*

32.7 KB

10.4 KB

the shell (was 41.7 KB)

yes

dist-*

31.8 KB

7.9 KB

(vendor)

—

apps-*

29.8 KB

9.2 KB

the game/app catalog data

data, not code

utils-*

27.6 KB

7.6 KB

shared utils

yes

VisualElement-*

26.4 KB

8.5 KB

framer-motion element core

partly — see §3

Gone from the critical path: esm-* (socket.io-client, 40.3 KB) and layout-* (68.2 KB, framer-motion’s layout projection).

Gone from the build entirely: icons-* (431.3 KB raw / 116.1 KB gzip / 91.3 KB brotli, 1,400 export specifiers) and lucide-react-* (157.1 KB, the barrel). What remains of lucide is createLucideIcon-* at 1.4 KB plus per-icon modules co-located into the route chunks that use them.

3 — What is still on the critical path and shouldn’t be

Both are open items, carried into 06-backlog.md:

  • schemas-* — zod, 71.0 KB raw / 16.6 KB brotli. Still there after nine route modules were fixed, because 246 client-reachable modules import zod — it is the shared schema layer, not a route-level accident. This is the one remaining critical-path item and it needs a decision rather than a patch: 02-critical-path.md §2 and 06-backlog.md §1.

  • VisualElement-* — framer-motion element core, 26.4 KB raw / 8.5 KB brotli. Reduced, not eliminated. Nine modules importing the full motion instead of LazyMotion’s m were fixed, which removed layout-* (68.2 KB — framer’s layout projection, and it had sat in the top five critical-path members for two audits unidentified). What still pulls VisualElement is a different and legitimate set of imports: Reorder + useDragControls (components/ui/sortable-list.tsx), LayoutGroup (components/user-builds/BuildGrid.tsx), useScroll/useTransform/useInView. Those are real framer APIs with no m-style lightweight variant, so removing them means putting their consumers behind a lazy boundary — a different kind of change. 06-backlog.md §12.

4 — Browser-level, and why to distrust the timings

Per-route, cold, unthrottled, all external origins aborted. Bytes are encoded transfer (CDP encodedDataLength), i.e. what actually crossed the wire after brotli/gzip:

Route

requests

JS

CSS

IMG

FONT

DOC

TOTAL

/

382

851–1062 KB

104 KB

6 KB

49 KB

60 KB

1134–1347 KB

/games

351

1105 KB

107 KB

0 KB

49 KB

89 KB

~1427 KB

/apps

365

1010 KB

104 KB

0 KB

49 KB

63 KB

~1306 KB

/library

334

873 KB

107 KB

6 KB

49 KB

61 KB

1159 KB

/news

342

855–1071 KB

104 KB

6 KB

49 KB

50 KB

1136–1352 KB

Request counts are stable to ±1 across runs and are the number to act on. Byte totals swing up to 20% run-to-run because the 3.5 s observation window catches a different tail of lazily-preloaded route chunks each time; ranges are given rather than false precision.

Corrected after publication. The IMG column originally read 296 KB for /games and 124 KB for /apps. Both were zero: the harness was classifying inline data:image/png;base64,… URIs as images because they match an image-extension regex, and CDP reports a nonzero encodedDataLength for them even though nothing crossed the network. /games and /apps fetch no raster art at all — the SSR HTML for /games contains zero <img> tags, and the cards are a CSS gradient plus a lucide icon. Filter data: URIs before attributing bytes.

Timings, same runs:

Route

FCP

LCP

long tasks

/

724–824 ms

1632–1924 ms

457–584 ms

/games

868 ms

2268 ms

798 ms

/apps

724 ms

2476 ms

612 ms

/library

716–852 ms

716–852 ms

303–435 ms

/news

644–660 ms

644–2184 ms

396–405 ms

Do not quote these as absolutes. This is a shared, CPU-contended container. Two first-touch samples came back at FCP 5768 ms (/library) and 11704 ms (/news) and were not reproducible on any subsequent run — cold route-module compilation, not a site defect. Anything under ~2× is noise here. What is meaningful is the same-harness, minutes-apart before/after:

Route

metric

before

after

change

/

requests

533

382

−28%

/

long tasks

813 ms

~570 ms

−30%

/

LCP

2060 ms

1904 ms

−8%

/

icon chunk fetched

yes

no

—

/games

requests

515

350

−32%

/games

long tasks

860 ms

697 ms

−19%

Byte caveat, stated plainly. The “before” run used performance.getEntriesByType('resource'), which the browser caps at 250 entries, so its byte totals were silently truncated and are not comparable to the CDP figures above. The byte win from removing the icon chunk is therefore asserted from the build (a 431.3 KB raw / 116.1 KB gzip chunk that was fetched on / — initiator trace in 02-critical-path.md §1 — no longer exists), not from a before/after browser diff. The request-count and long-task deltas above are same-harness and are comparable.

5 — Server-side

cold first request after boot:  1.55 s – 2.04 s TTFB   (three boots)
warm:                           26 ms – 80 ms TTFB

The cold cost is loadEntries() in startRequestResolver: TanStack Start loads every route entry module before serving any request, and there are ~855 of them (570 under app/routes/api/**). The SSR router bundle is 2.77 MB. This reproduces the 08-09 finding (1.84 s) unchanged, and is why a module-scope third-party client constructor is a site-wide availability risk rather than a feature-local one — see 05-server-edge-fonts.md §1.

server/nitro/warmup.ts masks this for the homepage after a deploy, so it is a cold-start concern rather than a live one — but every blue/green swap pays it, and any route the warmup does not prime pays it first.

6 — The harness

measure2.mjs, written to the repo root and deleted afterwards. Recreate it from this description; the important details are the three things that are easy to get wrong:

  1. Bytes from CDP Network.loadingFinished.encodedDataLength, not from performance.getEntriesByType('resource') — the latter is capped at 250 entries by default and will silently truncate a page like this one. If you do use the Performance API, call performance.setResourceTimingBufferSize(5000) in an init script first.

  2. Abort every non-origin request in page.route, and record the hosts you aborted. That both removes third-party network variance and produces the third-party inventory — which is how the nine Google Fonts routes were found.

  3. Attribute by URL extension, not by initiatorType. Modulepreloads and import()ed chunks report inconsistent initiator types across Chromium versions.

For chunk-graph analysis, walk .output/public/assets/*.js with two regexes — (?:from|import)\s*"\.\/([^"]+\.js)" for static edges and import\(\s*"\.\/([^"]+\.js)"\s*\) for dynamic ones — and BFS from the entry. Caveat learned the hard way: those regexes miss __vitePreload and route manifest references, so a chunk absent from the static closure can still be fetched on every page load. Confirm reachability in a browser (Network.requestWillBeSent + e.initiator.stack) before concluding a chunk is not on a page. That mistake nearly buried this audit’s largest finding.