[UC-9] Edge / mobile inference under power budgets

[w4ffl35]

see plans/use_case_edge_power_budgets.md — goal, reference architecture (train → prune/quantize → drift validation → NIR/bundle → InferenceSession → spikeforge-serve → clients → size/latency/energy /metrics), rate/delta coding, fc_small/sequence_mlp with pruning + activation/membrane quantization, acceptance (size ≤ budget, retained metric within 2%, p99 ≤ budget, energy estimate: true), MVP phases P0–P4, dependencies PT-W5/W1/W2/W3/W6/W7.

Spec: plans/use_case_edge_power_budgets.md

Refs #12 (umbrella tracking the ten chip-less production use cases).