bench(worldmodel): re-run the CEM-budget sweep on the repaired benchmark

[aurascoper]

Ledger node 16. Nodes 10 and 11 swept CEM knobs under the shared-generator defect, so their paired reading did not hold. Re-run with per-episode generators, a fixed episode set, n=200 and the continuous endpoint.

Node 11's conclusion survives. Its point estimate does not.

More CEM budget does not lift control — 0/5 seeds reach paired |t| = 2 on either endpoint in either mode. That part holds.

	default(3,64)	tuned(8,192)	delta
node 11 (as recorded)	0.342 ±0.151	0.325 ±0.144	−0.017
repaired, signal	0.335 ±0.011	0.360 ±0.016	+0.025
repaired, nuisance	0.337 ±0.023	0.354 ±0.025	+0.017

The ±0.15 node 11 called seed noise was mostly the benchmark's own defects — the spread is now 14× tighter — and the direction it reported was noise. A conclusion can be right while every number supporting it is not.

And there is a worse finding underneath

log(d_final/d_start) is positive in all 20 runs: +0.108/+0.082 (signal default/tuned), +0.128/+0.097 (nuisance). The planner ends 9–14% farther from the goal than it started.

Against trivial policies on the same episodes:

policy	signal	nuisance
zero-action	+0.034	+0.029
random-action	+0.046	+0.060
planner	+0.108	+0.128

The planner is worse than doing nothing. Paired per-episode, planner minus zero-action is positive in 6/6 mode × seed comparisons (+0.055 to +0.155, sign test p ≈ 0.03) — though only 1/6 individually clears |t| = 2, so this is a consistent direction rather than an individually significant effect. Same direction against random-action, 6/6.

Why this reframes nodes 11–15

They asked why the planner fails to beat chance. That framing is too generous. An uninformative planner scores like the zero policy. This one is consistently worse, which means something is systematically pointing it away from the goal.

That is a directional defect, not a power problem — and it is cheap to localise, which the last five nodes of sweeping were not.

Node 16 lists candidates in order and names the cheapest discriminating test — sign/axis convention between the latent displacement and the action applied in _step — rather than proposing another sweep. Note that node 14's one-frame remedy did not fix control, so the frame offset alone does not explain it.

Scope

Ledger-only change; no code. Verified: smoke test 2 consecutive runs, ledger gate and ADR gate green.