fix(worldmodel): pin the episode set, per-episode CEM noise, paired endpoint

[aurascoper]

Items 3–5 of the benchmark repair.

1. The episode set is a property of the benchmark, not the run

evaluate() passed seed = model_seed + 2 into _mpc_success, so every model seed drew a different episode set — "which episodes" was confounded with "which model". Between-set difficulty carried sd 0.0853, about half the observed baseline swing, and was being read as model variance.

_mpc_success now takes episode_seed (default 2) and drops seed entirely. Episodes, baseline draws and planner noise all derive from it and nothing else.

across 10 runs (2 latent dims × 5 seeds)	before	after
random_baseline_success	0.254 – 0.483	0.303 for every run
episode sets	5 distinct	1, byte-identical

2. A generator per episode

One torch.Generator consumed sequentially made the planning noise on episode i depend on how many draws episodes 0…i−1 had taken — a function of n_samples × cem_iters × horizon. Any arm touching a CEM knob got different noise on the same episode, silently defeating pairing even with the episode set fixed. Nothing in the panel would have shown it.

Nodes 10 and 11 swept exactly those knobs. That sweep was not the paired comparison it appeared to be — not necessarily wrong, but its "within seed noise" reading rested on an assumption that did not hold.

3. Paired continuous endpoint — log(d_final / d_start)

The ratio, not the difference: episodes differ enormously in how far the goal starts, so a raw difference is dominated by episode difficulty. The log makes "halved the distance" one effect size wherever it happens.

Measured, not simulated — paired ld32 vs ld64 on identical episodes and identical planner noise:

endpoint	mean |t|	median n for |t|=2
per-episode binary	0.62	~248
per-episode log-ratio	0.88	~154

1.43×, where a power simulation suggested nearer 3–4×. The mechanism the simulation identified is right — only 3–8 of 20 pairs are discordant, which is why binary is weak — but the size was overstated, and five seeds is a noisy estimate of it.

An earlier version of this measurement said 1.82×; that used an approximation of the paired binary sd from the two marginal rates. final_distance_per_episode is now returned so either endpoint can be computed exactly downstream, and the exact figure is 1.43×.

log_distance_ratio_mean is reported, NOT adopted as the adjudicator. Switching the primary endpoint in the same change that repins the episode set would confound the two. Adoption is a separate pre-registered decision that still has to clear the blind criterion.

Node 14's success_rate values no longer reproduce — by design

Per-episode generators change planner noise. The episode set for model-seed 0 is unchanged (the old effective seed was 2; the new fixed episode_seed is 2), but every seed now scores that same set.

Node 14's conclusions are untouched. The frame ratio, the Procrustes residual and the one-frame remedy were all measured within a run, never across seeds. Recorded in a dated Corrections to node 14, part 2 block rather than by editing node 14.

Verification

smoke test   3 consecutive runs, and at 1 and 12 threads
ledger gate  PASS (0 warnings)     ledger tests  OK
ADR gate     PASS                  ADR tests     OK
frame tests  OK (2 skipped under the shim, as designed)

Remaining

Raise n to ~200 — the last item, and now a config change rather than a design one.