fix(worldmodel): persist the EMA target encoder; pin the random baseline - #54

First two items of the benchmark repair, plus a correction to node 14 that the
work surfaced.

1. CHECKPOINT FORMAT v2 — persist target_encoder_state_dict.

v1 saved encoder_state_dict and predictor_state_dict only. The failure mode was
documented here, in frame_diagnostic.py and in ledger node 14 as "a restored
model has target == encoder, so it reports zero mismatch". That is WRONG, and
the truth is worse: __init__ deep-copies the target from the encoder AT
CONSTRUCTION, before any load, so a v1 restore leaves the target at its RANDOM
INITIALIZATION. Measured after loading encoder_state_dict into a fresh module:

  vs that module's own random init   0.00e+00   <- byte-identical
  vs the loaded online encoder       2.04
  vs the trained target              0.39

target == encoder holds only if you construct and never load, which is not what
using a checkpoint means. A v1-restored model plans toward goal latents from an
UNTRAINED network -- not the one-frame remedy arm, and not any condition that
was ever run. v2 restore verified exact (max parameter diff 0.00e+00).

This matters beyond tidiness: the registered acceptance criterion for the
benchmark repair says it must separate zero from random from MSE ON EXISTING
DATA. Run against v1 checkpoints that would have been scoring an untrained goal
encoder, and it would have looked like it worked.

train.py has no equivalent defect -- it saves model.state_dict(), which already
covers every submodule. export_coreml.py's GoalEncoder export is therefore also
unaffected; checked rather than assumed.

2. PIN THE RANDOM BASELINE — baseline_reps, default 100.

The baseline touches no encoder, no predictor and no CEM, only _step with
uniform actions, so precision costs env steps alone. It was estimated at one
draw per episode -- a scalar with binomial SE ~0.10 at n=20 -- and re-drawn per
seed, injecting its own sampling noise into every arm comparison as if it were
model variance.

Decomposing node 14's "0.15 to 0.45" swing, it is NOT episode-limited binomial
noise alone. Same episode set, varying only the draw: sd 0.0750 at reps=1,
0.0054 at reps=100, 0.0022 at reps=500. Different episode sets with reps pinned
at 100: sd 0.0853. So it is roughly half sampling noise and half genuine
episode-set difficulty. Reps removes the first half; decoupling the episode seed
from the model seed (next item) is required for the second.

Kept episode-paired rather than moved to an independent episode set: a paired
comparison needs P(random succeeds | THIS episode). Per-episode rates are
returned so a later paired analysis can use them as a difficulty covariate.
random_baseline_stderr is returned so a reader can see the pin held.

THE MODEL ARM IS BIT-IDENTICAL. Verified against node 14's recorded ld=32
success_rate at seeds 0-4 (0.30, 0.55, 0.20, 0.35, 0.35): exact match on all
five. starts/goal_pos are drawn before the loop and the CEM generator is
separate, so moving the baseline estimation after the model loop cannot
perturb it -- but that is an argument, and the check is the evidence.

3. Ledger gains a dated "Corrections to node 14" block rather than a retro-edit
of node 14 itself, matching the file's existing "Review clarifications"
precedent. Node 14's measurements stand: forward_eval.py has no torch.load and
trains in-process, so nothing it reported came from a restored checkpoint.

Verified: smoke test passes on three consecutive runs; ledger gate, ADR gate and
both test suites green.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DJJ8y1zo7WSqEmGNW8DQap