GENERAL NOTES

Method — one basis, two series. Every count on this sheet is read at build time from the checked-in plate www/atlas.lit-ui-router.dev/data/census-city.json, counted at origin/main @ 3f55e419 (2026-10-31): .ts/.tsx/.js/.jsx/.mjs under each member's source directory, excluding *.d.ts, *.test-d.ts, fixtures, dist/ and node_modules/; sloc is scc's Code count — string-aware, so the interior of a template literal counts as code rather than as one line. Sheet 3 throws test code away; this sheet does not — anything matching *.{spec,test,cy}.* or living under specs/ test/ tests/ __tests__/ cypress/ is counted into a second series and drawn as that member's annex. 52 members: 477 authored files and 42,967 sloc of source, plus 254 files and 46,674 sloc of spec. The packages district's 14 buildings reconcile exactly with sheet 3's source slab — 109 files, 10,708 lines — because both are queries over one census rather than two hand counts.

Every shipped building but ui-router-registry carries an annex bigger than itself. lit-ui-router: 3,262 authored lines against 9,354 of spec (2.9×). ui-router-server: 3,006 vs 6,856 (2.3×). lit-ui-router-mobx: 144 vs 579 (4.0×). eslint-plugin-lit-ui-router: 1481 vs 2560 (1.7×). lit-ui-router-effect: 285 vs 1093 (3.8×). lit-ui-router-ssr: 1485 vs 5085 (3.4×). navigation-location-plugin: 214 lines in 2 files under 7 spec files (5.0×). lit-ui-router-signals: 170 vs 855 (5.0×). ui-router-signals: 104 vs 185 (1.8×). ui-router-effect: 30 vs 62 (2.1×). ui-router-mobx: 117 vs 143 (1.2×). ui-router-registry is types only: 362 lines against a 63-line fixture (0.2×), because its tests are *.test-d.ts type assertions, which the census leaves out. The test budget lives exactly where the repo says it should — in packages/*, and nowhere else at that ratio.

Every road is citable. Solid roads are build edges: packages/* into the sample apps, the apps into docs, @tools/oxc-emit into all twelve published packages — each a workspace: dependency, each backed by turbo's build/docs tasks depending on ^build. Dashed soft roads are the test lane: @tools/dts-backtest depends on the runtime packages of the district and backtests their emitted d.ts against the TS 5.0 floor; @tools/lit-test-env and @tools/happy-dom run up into lit-ui-router's annex because they are its devDependencies and the harness its suites load; sample-app-lit-e2e depends on docs, and turbo's e2e depends on ^build and runs with @www/lit-ui-router.dev#wrangler:dev, so Cypress drives the built shopfront under its worker rather than a vite dev server. The accent road is the one that leaves a member twice — once from the source block and once from the spec annex — because the package's typecheck reads both: tsconfig.json includes src/**, specs and all (only the narrower typecheck:src excludes them), while the ^build:types dependency is what carries the result into the next district. One faint road stands for 24: that many members depend on @tools/shared, 17 of them instruments, and its src/globs.ts (#655) holds the custom-element glob three lanes share. Only the road to the largest is drawn, because drawing all 24 would turn the yard into hatching.

Severity in colour, mass in geometry — and they disagree. The tiers are sheet 3's, applied to whole members: @tools/release is solid red because it hosts published-diff, the only structure that halts a publish; red hatch marks the seven members that can stop a pull request — Cypress e2e, dts-backtest, compat-guards, lit-test-env, happy-dom, lint-elements, whose eslint lane runs at --max-warnings 0, and repo-checks, whose six checks — graph edges, task inputs, single version, patches, knip and dedupe — ride the lint lane; accent hatch marks @tools/workers-builds, which gates the docs deploy and nothing earlier. Here the two encodings happen to agree once, on release, and disagree everywhere else: the harness that can stop every pull request in the repo is 14 authored lines of @tools/happy-dom, sitting on the drawing's minimum footprint.

What the districts say. The product is 14 modest buildings — 10,708 lines — against 17,759 in the instrument yard and 4,927 in the proving ground. The broadest source block on the sheet is examples — 5,915 lines of stackblitz copies — with @tools/oxlint-anti-slop just behind at 3,665 and sample-app-shared, which ships to nobody and gates nothing, undefined at 2,582; two spec annexes outbuild examples. @tools/wintercg-globals is drawn on the floor with no mass at all: it is ambient types, and a census that hides its empty members is not a census.

Numbers by import, not by paste. Every count above and on the drawing is read from www/atlas.lit-ui-router.dev/data/census-city.json, the snapshot census-city.mjs writes out of the master per-file census; this module holds placement, tiers and prose only. The plate is measured at a named ref rather than at whatever a working tree happens to hold — counted at origin/main @ 3f55e419 (2026-10-31) — so every number here is citable to a committed file, and a member the sheet draws but the plate does not carry is a build error rather than a stale constant.