Event-driven runtime and energy accounting (WS-D)

Sparse runner

event_runtime.sparse_run() is a parallel, training-free inference path that propagates spike events instead of dense MACs. It returns a SparseResult with the same readout contract as the dense Trajectory, and dense_compare proves parity within tolerance. The dense path stays the untouched default.

SynapticCounter tallies SOP (synaptic ops), MAC (dense baseline), AC, and timesteps; for a sparse input SOP < MAC by the active-spike ratio.

Declared per-target cost tables

Each target carries a declared cost table under energy/costs/ (reference, norse, lava_loihi2, spinnaker2, speck, xylo), giving energy per SOP/MAC/AC and latency per timestep. Every table is "measured": false and carries its source.

spikeforge-energy

spikeforge-energy account --topology conv_net --target reference --sparse
spikeforge-energy report  --topology conv_net --target reference

account prints the report; report adds the sparse-vs-dense parity block. The energy_report WebSocket action and the benchmark --energy path carry the same payload; EnergyPanel renders it with a prominent estimate badge.

The estimate-vs-measured rule

Every report carries estimate: true and a basis ("declared cost table", or "unavailable" when a target has no table — never a fabricated number). Numbers are estimates from declared tables, never presented as measured, until a device reports its own timing.