The web layer derives nothing

The HTTP module reads through TraceView's five closed shapes and analysis, and serialises, filters and pages what they return. It computes no statistic of its own. Any number the interface needs that nobody derives yet becomes a new derived shape in analysis, with tests, and the command line gets it for free.

Considered options

Serving the interface from Next.js alone — API routes shelling out to the CLI and reading run directories — was rejected because the aggregates the mockups need (runs this month, spend against budget, per-tick trajectories per audience, sweep health) would then be written in TypeScript. FINAL_ARCH.md §5.10 records that this architecture exists because the previous design computed objection clusters in two places from the same verbatims with no stated owner: one concept, two implementations, guaranteed to diverge. Rebuilding that with a language boundary through the middle would make it harder to see, not easier.

Consequences

Two processes run locally instead of one, and every panel wanting an underived number requires a small analysis change before it can be drawn. That friction is the point: the interface cannot invent a statistic. Cross-run workspace totals read registry entries rather than scanning traces, which means a figure like "personas simulated" needs a field on the entry rather than a walk over every partition.