fix: store z_mu instead of sampled posterior in latent precompute

[Can-Zhao]

encode_stage_2_inputs() returns z_mu + sigma*eps. With weak KL regularisation (kl_weight~1e-7), the noise term accounts for ~27% of latent variance and acts as an irreducible floor on the diffusion L1 objective. The released checkpoints were trained on z_mu (encode()[0]); this aligns the script with that practice.

Fixes #45.

[greptile-apps]

Confidence Score: 5/5

[High risk] Changes how training data embeddings are computed and stored.

The PR appears safe to merge; no outstanding findings remain.

Summary

The PR changes latent precomputation to store the autoencoder’s posterior mean instead of a sampled latent, and updates the embedding instructions to match.

• The README now prominently advises regenerating embeddings created before the fix.

_{Reviews (3) · Last reviewed commit: "fix: split long line in data/README.md t..."}