feat: open forward model (physical imagination) for the 85 mm body

[xebastian153]

Closes the proposal in #6, in the shape you asked for: standalone runner + weights first, planner split into its own file, at policy/forward/ beside the walk policy.

What this adds

policy/forward/

• growbot_forward.js — the runner. Pure JS, no deps, browser or Node. History ring, one-tick step(), imagine(plan) open-loop rollout. 82 lines.
• forward_85mm.json — the weights. 24,841 params (128×2, swish), same size class as policy_85mm.json.
• growbot_planner.js — optional planToMatch(fwd, targetTrace): a small CEM over an action chunk, the mimic game. Separate so ports and the harness can take the runner alone.
• test_forward.mjs — verifies the JS against the trained PyTorch net: single step 7.5e-6, 25-tick rollout 5.0e-6; and that the planner beats hold-still on a reachable target. node test_forward.mjs → PASS.
• README.md — contract, usage, measurements, caveats. Same layout as policy/README.md.

Contract

Mirrors the walk policy's so they share one calibration: last 5 (imu6, [aRight, aLeft]) pairs newest first, same IMU frame and signs you feed GrowBotPolicy, actions in radians of leg swing. Output is the IMU change over one 20 ms tick; angles are (sin, cos) internally so yaw wrap and a roll through ±π never jump.

Measured in the twin (policy/Harsh_policies/DR_RMA_EXPORT)

Imagined roll/pitch within 0.2 rad of truth, held-out episodes:

horizon	nothing-changes	linear	this model
100 ms	85.9 %	93.5 %	95.9 %
500 ms	59.0 %	75.0 %	83.6 %

Gains concentrate where the video locates the gap: fast motion 41 → 86 %, tipping/fallen 58 → 89 %.

Mimic game, 40 held-out 2 s traces, plan through imagination and execute in the twin: hold-still 0.210 rad RMSE, planning without a forward model 0.220 (worse than doing nothing), with this model 0.095, beating hold-still on 39/40. Replanning every 100 ms is the optimum.

What it isn't

Sim-only — it has learned MuJoCo. The on-device error-driven update (the learning half) is the follow-up, and I'll take that to #world-model. Nothing here touches firmware, protocol or the harness.

Source, data generation, evaluation: https://github.com/xebastian153/growbot-cerebellum

Checklist

• One change, one folder, 251 lines + weights JSON
• No secrets, keys or env files
• Test runs green from inside the fork
• License: PolyForm NC, same as the walk policy, weights included