feat(worldmodel): action-conditioned probe — the action is not represented

[aurascoper]

Node 17's test (c): measure how much of which action was taken survives into the representation. The prediction was too generous.

Result

features	ax	ay	mode_bit
TRUE (pre, post) — ceiling	0.7096	1.0000	−0.055 ✓
LATENT (z_pre, z_post)	−0.1676	−0.1199	−0.260 ✓
latent delta	−0.1174	−0.0963	−0.150
z_pre only — control	−0.1183	−0.0845 ✓	−0.156

ay is exactly recoverable from the true states — ay = (pos' − pos − vel'·DT)/(0.5·DT) is linear, so R² = 1.000 — and not at all from the latent. Negative R² is worse than predicting the mean. The representation carries nothing a linear probe can use. Not the ~6% node 17 implied; zero.

Both controls behave: mode_bit is action[2], which _step never reads, and z_pre alone cannot see the transition.

The mechanism, and the sharpest number in the ledger

peak_amp 0.9971   chi 0.9498   offset 0.6856   pos 0.6537   vel 0.0326

The three near-perfect factors are per-trajectory constants. The two that evolve are the two worst — and vel, the channel ax acts through, is at zero.

So ax is unrecoverable in principle: recovering it needs vel and vel', and neither is encoded.

For pos the arithmetic is just as decisive: R² 0.6537 against sd 0.5812 leaves encoder residual 0.342, while the one-step action effect is 0.120 — SNR 0.35. The action's fingerprint is a third the size of the encoder's own noise on the only variable it can move.

Adopted into the panel, with its control asserted

action_recovery now ships in the panel, and smoke_test asserts mode_bit <= 0.05.

Mutation-checked: leaking the probe's target makes mode_bit score 1.000 and the assertion fires. That control is the only way this metric can silently lie, so it is asserted rather than eyeballed.

Corrects node 17

"Under-represents by ~3.5×" becomes "does not represent at all". Node 17 measured the predictor's response to an action it receives as an explicit input; this measures whether the encoded transition retains which action was taken. The second is the load-bearing quantity and it is zero — which also means node 17's 6.1% predictor response is fitting noise, not a weak signal.

This closes the nodes 6–17 arc

Control was never reachable from any representation-side change, because the encoder discards the action's effect and encodes the frozen factors nearly perfectly instead. Every panel metric that improved was measuring the constants.

vel at R² = 0.033 has been visible in every run since node 3 and was never read as a control result.

Next test, and it may retire the line

Check whether vel is present in _render_state at all. If the observation does not carry velocity, the task is unobservable rather than hard — a generator defect, and it would retire the control line as specified rather than motivate another encoder change.

Verified: smoke 3 consecutive runs; ledger, ADR and test suites green.