morel — Limitations

This document enumerates known boundaries of the implementation.

Out of scope (v0.1.0)

Distributed training — single-process; multi-GPU via DDP is a future extension.
Streaming ingestion — the data pipeline assumes the full dataset fits in memory for the k-core and graph construction steps.
Real-time fine-tuning — the serve stack is for inference only.
Modality auto-detection — modality names and dimensions are configured; there is no inference of modality presence from data.
Multilingual sentence encoders — only sentence-transformers is supported; custom raw-modality encoders plug in via the FeatureEncoder Protocol.

See FIDELITY.md. Notable:

MAGE — best-improvement hill climbing + sorted boundary (vs. paper's ambiguous "greedy").
Joint encoding — learned linear-pool attention vs. paper's vague "attention over the entire set".
Top-P — implemented as Top-K (post-softmax topk + renorm); same numerics in expectation.

Single-node graphs: PE is all zeros; downstream transformer handles via mask broadcasting.
Empty adjacency: warnings logged; downstream receives a zero PE.
Disconnected anchors: ACS logs a warning and returns either the anchor set or an empty set, depending on Config.retrieve.acs_fallback.

CUDA non-determinism in torch.sparse.mm requires explicit torch.use_deterministic_algorithms(True) if exact bitwise reproducibility is required on GPU.
The fixture cudnn.deterministic = True may slow down training on GPU.
Some torch versions do not roundtrip set_rng_state bytewise; the state() snapshot includes the RNG state but exact reproduction may require re-running with the same torch version.