GENERAL NOTES

Method — one meter per member, nothing installed by hand, nothing left behind. Footprints, annexes and districts are sheet 7's own plate (census-city.json); the light is census-shadow.json, metered at origin/main @ 3f55e419 (2026-10-31), so meter and census read one tree at one ref. 37 of the 52 members carry a meter. The probe never edits a repo file: members that declare test:coverage are metered by the tree's own unmodified turbo run test:coverage; node:test members by their own test script re-run with --experimental-test-coverage and the lcov reporter; the vitest members by their own script with --coverage.enabled --coverage.provider=v8. Each meter reports lines on its own basis (v8 remaps to executable lines; node counts raw lines), so brightness percentages are per-meter and are never summed across meters in the drawing — only the schedule's grand total does, and says so. LIT means a source file the suite actually executed; extent is those files' sloc over the member's src sloc, both from the same ref. One honest wobble found by re-running the probe four times over: node --test's branch denominator for №31 came back 213 on two runs and 215 on two more, with 191 hit either way — a 0.9-point swing on one member's branch figure and nothing else in the plate moved. Branch discovery under V8 is not perfectly repeatable, and this plate says so rather than pretending the last run is the only one.

The product glows wall to wall — the budget lands where the house says. The published packages meter 99.7% of 9,921 lines over 96.8% of their metered source. What little stays dark is entry barrels and interface.ts — type declarations, which no runtime meter can light. ui-router-registry stands dark for the same reason: it ships types only, has no test script, and its type tests run under typecheck, where no meter reads. This is the priority made visible: library coverage outranks docs coverage, and the annexes sheet 7 drew at 1.2–8.0× their buildings turn out to buy near-total light.

The yard's habit: bright cores, dark wrappers. The instruments repeat one pattern — @tools/release, build_and_test, workers-builds, lcov-rebase: the .core.ts logic is lit and the command-line file in front of it is not. 4,270 sloc across 72 files of process-edge code is never loaded by any suite, while the logic behind it meters 94.7%. compat-guards is the pattern at its extreme: only ranges.ts is unit-lit; its guard lanes run for real in CI, where no meter follows. @tools/happy-dom splits its light: its spec is a conformance canary pointed at happy-dom upstream, and it lights the one workaround it guards, inner-html.ts, while append.ts is lit only from lit-ui-router's lamp, as borrowed light.

What the meter cannot say, the plate refuses to fake. The two demo apps and the Cypress host are hatched in accent, not banded: e2e light is real — the rig drives the built docs site — but no lcov leaves it, so it is drawn as light of unknowable brightness and labelled unmetered. That judgement is the only one in the probe, and it is verified rather than assumed: the run throws if the rig has stopped being a Cypress suite. sample-app-shared runs its unit tests green, yet browser-mode vitest cannot fetch a coverage provider the repo never installed; its light is an outline, and the probe proves it by running the suite a second time without the meter. The members that are dark are dark because nothing tests them, and two of them (dts-backtest, lit-test-env) spend their working lives inside other members' suites.