Metabolism: decision-cadence (think slower than you move) + empty-block skip; flat typed-array bytecode; WebGPU decision-layer slice (experimental) + heavy-world benchmark

metabolism/EVOLUTION.md

@@ -249,6 +249,53 @@ and wander); rock has a budget that survives a whole film (calm songs keep
 sparse stone, percussive ones accumulate walls); agriculture consumes stone
 instead of minting it. 44 tests green; realtime holds at 640×360.
 
+### The decision layer on the GPU (analysis, first slice landed)
+
+The profile of a heavy frame (field 128, ~40 000 organisms planted) says:
+
+| layer | share |
+|---|---|
+| the strain VMs (`runProgram`) | ~35% |
+| senses/gradients/action plumbing (`runStrains`) | ~20% |
+| field passes (block sums, reaction, `substep`) | ~16% |
+| advection (+allegiance, burden) | ~7% |
+| camera + render | ~10% |
+
+Three moves, in the order they pay:
+
+1. **Think slower than you move (landed).** The programs ran at 150 Hz —
+   every substep — though a mind's decisions barely change in 1/30 s. The
+   world now thinks every `vmEvery`-th substep (default 2: ~15 thoughts/s)
+   and coasts on its last decision between thoughts; blocks with no
+   population think not at all. Measured on the heavy world
+   (41 k planted, sim only): 25.0 → 20.3 ms/frame at ×2, 17.7 at ×3 — with
+   statistically identical ecologies (38.6 k vs 38.7 k final). The tests
+   pinned the rest.
+2. **The GPU (first slice landed, experimental).** The decision layer is
+   ideal GPU work: m uniform programs over ~4 000 independent blocks —
+   lockstep, only the data differs. `src/sim/gpu.js` holds a WebGPU compute
+   path for exactly that piece: senses/grads packed on the CPU, the flat
+   bytecode uploaded per strain, registers resident on the device, one
+   dispatch per strain, the action buffer read back per substep.
+   `tools/gpu-check.js` (browser-side) validates the GPU result against the
+   CPU interpreter and times both. Status: written against the WebGPU API,
+   awaiting a device to run on; NOT yet wired into `World.step` — that
+   wiring needs an async stepper (the readback is asynchronous), which the
+   demo can adopt and the exporter need not.
+3. **The rest of the field** (block sums, advection, reaction) is the next
+   and larger GPU slice — element-wise and stencil passes over the fields,
+   with the population/politics staying on the CPU. That is what unlocks
+   field 256 and fatter casts.
+
+**Determinism policy.** GPU float ordering differs from the CPU's, so a
+GPU-accelerated preview may drift slightly from a CPU-exported film. The
+policy: the prepared-artifact/export pipeline stays CPU and remains the
+canonical pixel source; the GPU path is a preview accelerator whose world
+is the same world to within float noise. If preview/export bit-identity is
+ever demanded, the exporter must run on the same backend as the preview —
+a node WebGPU binding (e.g. a wgpu native addon) is the door for that, and
+it is explicitly out of scope until the film needs it.
+
 ### M3 — the lens: travel, don't cut
 Replace the cut-to-a-point grammar with a **move grammar**: *track* (follow a
 moving front), *reveal* (pan along a coastline), *establish* (hold wide and