Property engine: change propagation (deferred) #220

Deferred. Opened so the design and its trigger are recorded.

Why not now

Constant folding is the only incremental part of update(); everything else runs every time. In the FFT solver the strain changes at every voxel and every iteration, and each voxel loads its own history, so nearly everything would rerun anyway. Change propagation pays off only in hosts that hold the strain fixed while something else changes (temperature-only updates, parameter studies). No current host does that.

Design, when needed

• Every property gets a version number. A write bumps it: external sources on bind(), outputs when their callback runs, history on commit(), revert() and load_history().
• Each callback stores the versions of its declared reads from its last run. It is skipped when they are all unchanged and it is not stateful().
• This relies on the declared reads (Property engine: graph edges per callback from the declared reads #169–One declaration style: deprecate the old callback declarations, author guide (#172) #183), which the debug read check proves complete. Callbacks without a declaration always run.
• Constant folding becomes a special case: a value whose version never changes after its first run. The single global revision counter turns into per-property versions, which also covers Property engine: invalidate on the write, not on get_mutable() #186.

Risk

A skipped callback that has side effects, or reads something it did not declare, gives stale values silently. The read check catches undeclared reads only in debug builds.

Trigger

The first host that updates with the strain held fixed.