Operations
This section is about the unglamorous half of a shipped app: what it costs to start, what it tells you when something goes wrong, and how it behaves on each target. None of it is required to build a first screen, and all of it is what decides whether a bug report is a five-minute fix or a week.
Three tools exist for it, and each is measured rather than asserted:
bun scripts/verify.tsboots a real window on the real IPC path and asserts the security policy, the panic boundary, the storage round-trip and the multi-window contract. It is the source of the guarantees claimed elsewhere in this book.bun scripts/bench.tsmeasures cold start against a barewry+taobinary, and with--dispatchmeasures the async dispatch path.voussoir dev --verboserecords reload, dispatch and denial events while you work.
What "verified" means here
A claim in this book is either measured by one of those tools or it is not made. The
table in Platform support marks a cell cargo check when the code
compiles for the target but no window was opened, and says so rather than implying
runtime coverage.
In this section
- Diagnostics — the
--verbosemode and what it records. - Benchmarks — the numbers and how to reproduce them.
Coverage
./scripts/coverage.sh reports line coverage for the library's own tests and
fails under a floor. The baseline is 67.58%, measured with the default
features and with test and fuzz modules excluded, so the number describes the
shell rather than its assertions. The floor sits just under it: a gate set at the
exact current number fails on honest changes and gets deleted, while one a point
below still catches code added without a test.
It uses the llvm-profdata and llvm-cov the toolchain already links against,
so there is nothing to install and CI runs the same script. Raise the floor when a
release closes a real gap; never lower it to make a build pass.