fix(optim): ship the CMA-ES distribution mean, not the noisy xbest (#69)

[minion1227]

Fixes #69

Re-submitted from my personal sn74-minion1227-69 branch per CONTRIBUTOIN.md (previous PR #85 used a non-conforming branch name).

Problem

SepCMAES.best() returned pycma’s result.xbest — the single candidate with the best evaluated fitness across the run. TRINITY’s training fitness J(θ) = E[R(τ)] is a stochastic signal: each candidate is scored on a freshly sampled minibatch with a sampled policy (sample=True). Under a noisy objective, xbest is selection-on-noise — the luckiest evaluation, not the policy with the best expected reward. Since train.py saves best()[0] as best_theta.npy, the shipped router was chosen by minibatch luck and overfits sampling noise (the same noise the README notes swung random routing 0.792 → 0.733 between runs).

Fix

Return the distribution mean (result.xfavorite) as the solution vector. pycma documents xfavorite as the recommended estimate of the optimum for noisy problems, and it is the maximum-likelihood point of the search distribution. This is the issue author’s first proposed option ("ship the mean"), kept minimal and self-contained.

Because train.py already saves whatever best() returns, no change to train.py is needed — best_theta.npy now holds the mean automatically.

Kept stable on purpose

• The reported fitness stays the best observed training fitness (self._best_f), so run()’s history["best_fitness"] and the S7 smoke contract remain monotone; only the shipped vector changes.
• The "no evaluation yet" RuntimeError before the first tell() is preserved.
• Falls back to the locally tracked incumbent if a backend has no xfavorite.

Tests

Adds tests/test_sep_cmaes.py (CPU-only, torch-free, importorskip("cma")):

• best() returns xfavorite, not xbest
• best() raises before the first tell()
• reported fitness is the running max of observed fitness
• run()’s logged best_fitness stays non-decreasing (S7 regression guard)

Full root suite: 165 passed. ruff check clean on the changed files.