feat(worldmodel): additive task channels — representation repaired, control unmoved

[aurascoper]

Node 20's structural candidate. c0 = alpha·(1 + 0.10·pos) → c0 = 1.0 + pos, likewise c1/vel. _step untouched.

No tunable constant. JEPATransitionDataset z-scores every feature on train statistics, so any gain divides straight back out; the 1.0 is a DC offset for positivity that the same z-scoring removes. This restores _render_state's own stated intent — its comment already claimed these two channels carry the task factors.

Gates — registered before implementing, including G5

gate	result
G1 raw probe	pos 1.0000, vel 1.0000	✅
G2 oracle ladder	100.0 / 91.5 / 73.5 / 37.0 / 34.5 — bit-identical	✅ dynamics untouched
G3 alphaPower post|pre	1.000, unchanged	declared in advance as not addressed
G4 chi recoverable	0.6849	✅
G5 (the gate node 20 lacked)	below	✅

G5 — factor recovery under the incumbent objective, 3 seeds

factor	node 20 baseline	after
pos	0.6065	0.9824	task
vel	0.0227	0.9740	task
chi	0.9566	0.7660	nuisance
peak_amp	0.9947	0.8859	nuisance

The encoder inverted its priorities exactly as intended — task up, nuisance down. vel went from unencoded to near-perfect.

Action recovery transformed

ax −0.1676 → +0.8309, ay −0.1199 → +0.3709, mode_bit still −0.2807 so the negative control holds. Node 18's "no recoverable action information" is fully resolved.

The frame offset also improved 16× without being targeted: node 14's A/B ratio 70.6 → 4.25.

And control did not move

	before	after
mpc_success	0.35	0.353 ±0.034 (random 0.361)
log-ratio	+0.0886	+0.0582 (zero-action +0.0341)
planner correlation	+0.0375	−0.0206
sign agreement	45.5%	49.0%
|produced| / |needed|	33%	31%

Still at chance, still worse than doing nothing. The planner remains uninformative on a representation that now contains what it needs.

Verdict

Keep the repair — it passed every gate and fixed a real, measured defect. But nodes 17–19's diagnosis is necessary but not sufficient.

This is the third defect found, fixed, and followed by no change in control: the frame offset (node 14), the benchmark's resolving power (node 15), and now the action channel (nodes 17–19). Each was real. None was binding. That pattern is itself the finding — the control failure is not one broken component.

Next suspect, with evidence

Rollout error compounds 11.3× to 17.4× from h=1 to h=8 while horizon=6:

seed 0   1-step 0.00103   h=8 0.01734   17.4x
seed 1   1-step 0.00161   h=8 0.02302   14.8x
seed 2   1-step 0.00164   h=8 0.02898   11.3x

A CEM plan feeds the predictor its own output for six steps, so the cost being minimised may be computed on a state that does not match the trajectory executed.

Test named: score the CEM against a rollout through the true env on the same episodes and plans. If control jumps, the predictor's rollout is the bottleneck; if not, the fault is in the cost or the goal latent, not the forward model.

Verified: generator and forward_eval smoke tests pass; ledger and ADR gates green.