Instrument your fw-02/fw-05 runtime to display: component tree, render counts, state changes,
network requests, performance marks.
Do not perturb what you measure. Verify: no strong references retained to component instances
(a bi-04 leak), no observable state writes that trigger re-renders, overhead low enough that the
timings remain true. Measure the overhead and state it.
Deliberately build the leaking version first — retain every instance — and find it in a heap
snapshot.
Study the Chrome DevTools Protocol. Confirm that DevTools is a protocol client, and that
automation uses the same protocol.
Put one thing in CI. Pick a signal you currently inspect manually — trace-derived long tasks,
CSS coverage, an accessibility-tree assertion, or composited layer count — and wire it into a
pipeline as a regression guard.