FediAID artifact update: reproducibility improvements and five-seed MMR sensitivity

This release addresses artifact-review feedback and extends the optional MMR sensitivity analysis.

• Reproduction environment: pinned the validated core runtime and tokenizer dependencies and the DeBERTa encoder revision; added SentencePiece and the protobuf version constraint.
• Stricter verification: both evaluation workflows now reject missing references, incomplete coverage, duplicate results, non-finite metrics, and invalid tolerances.
• Clearer evaluation instructions: documented a complete raw-text reproduction route, aligned raw and embedded dataset names for reference comparison, and corrected command examples.
• Per-seed references: added expected F1/AUC values for all 60 dataset/seed combinations in the main evaluation.
• Five-seed sensitivity analysis: extended the optional sweep from seed 0 to seeds 0–4. Each platform is averaged over five seeds before computing the mean and population standard deviation across platforms.
• Documentation and licensing: documented the released checkpoint’s training data, split construction, settings, and training command; added an MIT source-code license.

The released checkpoint, memory bank, main expected metrics, and external evaluation configuration are unchanged.

On-the-Fly Centroid Normalization Alignment; Precomputed Reproduction Unchanged

Patch release correcting centroid normalization for the raw-text/on-the-fly embedding path. Post embeddings are now normalized before being averaged into retrieval centroids.

This change does not affect the precomputed-embedding path or the documented reproduction workflow, which continues to use the released precomputed embeddings.

No changes are required for reproducing the reported results using the documented procedure.

S&P 2027 Artifact Evaluation Snapshot v1.0

Artifact snapshot submitted for the IEEE Symposium on Security and Privacy 2027 Artifact Evaluation.

This release contains the evaluated version of the FediAID artifact.