GENERAL NOTES

Method: a generateBundle census on a real production vite build of apps/sample-app-lit-vanilla, run inside a materialized, installed archive of the measured ref rather than the working tree — apps/sample-app-lit-vanilla · counted at origin/main @ eb1162da (2026-10-18). Every rendered module is attributed to its package by an explicit grouping table whose misses land in a loud other row (this print: none), and every footprint, height, schedule row and door price is read from the checked-in plates at build time — a group the drawing places that the plate does not carry is a build error, not a stale constant. Footprint is post-tree-shake source kept; height is generateBundle: module renderedLength scaled by its chunk gz/rendered ratio — the honest wire share. Beside it, the repo's own codecov bundle-analysis probe (tools/bundle-probe, the <pkg>-<label>-esm series CI uploads) prices each exported entry at the same ref, read here from www/atlas.lit-ui-router.dev/data/census-doors.json. This closes the survey as a quartet: the source (7), what npm delivers (8), what the wire carries (9) — and here, who occupies those bytes. Sheet 11 is this sheet's split view: the same probe cut by published package, every entry priced alone. Whole bundle: 691,859 kept → 386,053 emitted → 125,269 gz in 17 chunks.

Reconciliation with sheet 9. The two sheets do not meet at an identity, and the reason is a rule: on sheet 9 a chunk counts where it is first claimed, so ui-router-visualizer.esm sits inside app: vanilla rather than beside it. Straight from the two plates: sheet 9's vanilla district is 130,185 gz over 18 files; this bundle is 125,269 gz over 17 chunks; the residual is 4,916 gz over 1 file. That file is the one emitted asset in the app's output that is not a chunk — the api-viewer custom-elements manifest, JSON a module census structurally cannot see, since it walks bundle entries of type chunk. The app's stylesheet is not in the gap: sheet 9's pattern table files .css under site css before the app districts are reached. So 125,269 + 4,916 = 130,185, and the named residual is a manifest, not a rounding error.

The library is a skin over the machine. The biggest building in the bundle is not the router package — it is @uirouter/core at 26.0 KB gz, 21.3% of the wire, 64 modules. lit-ui-router itself is the small accent building: 7.2 KB, 5.9%, 15 modules. The two independent measurements reconcile: the codecov door prices the bare . entry at 8,957 gz with deps external, and this app's census pulls 7,354 of it — the 18% difference is what the app's own imports leave on the shelf.

The chrome outweighs the machine. marked, dompurify and @api-viewer total 27.2 KB gz — more than the routing machine they document — and the prebundled @uirouter/visualizer adds 24.2 KB on its own, opaque to tree-shaking because it arrives already bundled. None of it is on the eager path: the panel renders behind a feature flag and rides a lazy api-docs chunk. This census counts who occupies the bundle, not when they arrive; sheet 9 draws the critical path. One lit major serves all of it — 8.7 KB gz across 19 modules, no second copy riding in with the docs-viewer stack, because a scoped pnpm override (^3.3.3, a floor, not a pin) holds @api-viewer/* and lit-dialog to it.

The app's own district is the small one. lodash-es is 7.5 KB gz from 135 modules and 65.9 KB of kept source — isEqual's _base* family, the largest module count in the bundle and its own chunk, which is what four imports truly cost once tree-shaking has done its work. Beside it, the app's own source is 2.4 KB and sample-app-routes — the route table the sample apps and the docs site import — is 332 bytes kept, 66 on the wire, one module: the smallest building on the map.

The runtime tax is near-zero. All bundler machinery — oxc decorator helpers, vite's preload helper and polyfill, a stray tslib — totals 5.4 KB kept, 1.2 KB gz, 0.8% of the kept source, in 4 modules. No duplicated-helper problem.