A screen shows the latest reading of its sources at once, and never waits for a handler

Rationale

On 2026-10-12 the owner called the UI "tragically slow" a day after it became the primary surface. The measurements were plain. /ui took 2.2 to 2.5 s on every request, because health.report runs probes and was part of the first paint. The first request after any change of the repository took 5.6 to 8.1 s on three screens, because a cache keyed by the repository's state is empty exactly then. One panel took 75.7 s. A cache had been added the same day and worked wherever it applied.

The defect was not the cache. A page computed while a person waited, and a page costs its slowest source. project.cache-is-invisible rightly forbids a cache from answering with a value from another state of the repository, so no cache could close the gap. ADR 0142 takes the computation off the request instead.

Required behaviour

1. A screen asks for readings. Every source of a screen is read through ui::source::ask, ask_all or ask_input, which ask Context::read. Outside ui/source.rs and tests, nothing under apps/majordomus-cli/src/ui calls execute.
2. A known source costs a lookup. A source read in the repository as it is now is handed over from memory. Its handler runs on a worker of capability::readings, behind the caller.
3. A moved repository is not waited for. When the repository has moved, a caller waits for a new reading no longer than its patience (50 ms for a page) and is then handed the reading it has, marked stale.
4. A reload is not waited for. A screen is drawn from the generation the server has (Live::view_without_waiting). When the repository has moved past it, the screen says so and the next generation loads behind the reader.
5. A reading says how it stands. A stale or pending reading is shown as that, and a stale one with its age. Its source line carries data-catching-up until a current reading replaces it. A current reading older than five seconds says its age.
6. An answer stays exact. execute runs the handler or returns what the handler would return. Nothing a reading does changes what the API, MCP or the command line answer.
7. It can be switched off. MAJORDOMUS_CAPABILITY_CACHE=off makes a reading run its handler and wait for it.

Failure behaviour

test/cases/1096_no_screen_waits_for_a_source.sh fails when a file under apps/majordomus-cli/src/ui other than source.rs calls execute outside a test module, when a served screen waits for its sources after the repository changed, or when a screen that is catching up does not say so. The readings tests in apps/majordomus-cli/tests/executor.rs fail when a known source is waited for, when a moved repository is waited for past the patience, when two callers of one source run two handlers, or when a failing or panicking source leaves no reading.

Verification

MJ_TEST_JOBS=2 bash test/run.sh 1096_no_screen_waits_for_a_source
cd apps/majordomus-cli && cargo test --test executor

The latency budgets a screen is held to, and the gate that measures them, are I2489.