Compose performance

Core principle

Measure one user-visible transition, identify the runtime axis that causes the work, then apply the smallest correction at the phase or boundary where that axis begins.

Procedure

1. Reproduce one concrete transition and capture the observable evidence: recomposition counts, compiler reports, profiler data, or a clear trace.
2. Classify the primary axis: parameter stability and skipping, State read phase, or snapshot state written back into an earlier phase.
3. Check for a false lead: a real data change, a correctness defect, or an unchanged lazy item that is expected to recompose.
4. Read the corresponding focused reference before proposing a change.
5. For review work, pair every finding with the smallest evidence-supported repair. For a false stability promise, explicitly say to replace mutable non-snapshot properties with immutable data or snapshot-observable state, verify that contract, and only then decide whether an annotation remains needed. For a phase problem, name the layout or draw consumer where the changing read or calculation should move. Do not stop at diagnosis.
6. Change one axis at a time and re-measure the same transition.
7. Finish when the evidence improves at the observed boundary without hiding state changes, caching stale values, or moving work to a less correct owner.

Topic router

RED/GREEN agent scenarios

1. RED blames unstable parameters for unchanged lazy rows that recompose during a focus transition. GREEN checks composition and layout back-writing first.
2. Novel case: an animation value controls only drawing. RED identifies the composition read but stops at diagnosis. GREEN moves the State read and its geometry calculation into the draw or layout consumer.
3. Counterexample: a screen visibly recomposes because its displayed model actually changed. GREEN does not add stability wrappers or caches merely to lower a count.