Rename Bekko Static Embedding v1 and replace results with FP16 inference

Migrate the results to the model's new official Hugging Face name and correct
inference precision. Delete the 563 existing files under
hakari-results/hotchpotch__bekko-static-embedding-multilingual-v1/ and add 563
newly evaluated files under hakari-results/hotchpotch__bekko-static-embedding-v1/.
This supersedes the FP32-compute evaluation merged in PR #40.

The model was renamed on 2026-09-14 and remains multilingual. The new run pins
hotchpotch/bekko-static-embedding-v1 revision
76c36707a7814a747914ed05168420696a8123d0 and uses --dtype fp16 for model
inference, with actual model parameters and encoded outputs verified as
torch.float16 at all six supported widths. The old result JSON was not relabeled
or reused: every standard task was reevaluated under the new identity.

Default Overall micro nDCG@10 x 100: 45.88383315, over 550 canonical tasks.
Overall candidate-reranking micro nDCG@10 x 100: 47.52831648.

Coverage is complete: 563 standard tasks, 30 conditions per task, and 16,890
result rows across tasks and conditions. Every condition has 550 canonical Overall tasks after the
normal duplicate exclusions. Conditions include base, all five truncations,
full-width int8/binary search and rescore, and every truncation x quantization/
rescore combination. All aggregate scores are finite and all retrieval and
candidate-rerank top-100 query sets are consistent across conditions.

Model card: https://huggingface.co/hotchpotch/bekko-static-embedding-v1/blob/76c36707a7814a747914ed05168420696a8123d0/README.md

Summary

Per-component nDCG@10

The generated template groups MNanoBEIR by task name (381 grouped units from
550 raw tasks), so its grouped summary differs from default Overall micro.

Reproducibility

Reproduction command

Equivalent full-run command, using a fresh results directory:

OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 TOKENIZERS_PARALLELISM=false \
uv run --with sentence-transformers==5.6.0 --with transformers==5.14.1 \
  hakari-bench evaluate dense \
  --model hotchpotch/bekko-static-embedding-v1 \
  --model-revision 76c36707a7814a747914ed05168420696a8123d0 \
  --all --dtype fp16 --device cuda:0 --retrieval-score-device cpu \
  --batch-size 4096 \
  --embedding-variant truncate:512,256,128,64,32 \
  --results-dir output/bekko-static-v1-fp16-20260914

The actual run validated NanoBEIR-en/NanoArguAna first, then partitioned the
remaining exact standard task manifest across 12 workers on cuda:0/cuda:1.
Each shard replaced --all with explicit --dataset and --split values and
pinned --dataset-revision to the revision recorded by the preceding run.
Every task's recorded dataset revision was audited against that run. No result
from the previously interrupted old-name FP16 run was reused.

Runtime and validation notes

• Standard PyTorch/SentenceTransformers StaticEmbedding loader; no ONNX or
  OpenVINO backend and no custom remote code.

• Stored weights are FP16, and model inference/encoded outputs are FP16.
  The checkpoint weight SHA-256 is
  586e165a868f306822174e8110020fa8759577cc62a537a702941747732858f8.
  It matches the prior FP16 delivery; the new repository revision also includes
  the rename and optional backend artifacts.

• Standard HAKARI post-encoding scoring retains its existing FP32 conversion
  path on CPU. This is a --dtype fp16 model run, unlike PR #40's FP32 model run.

• Empty checkpoint prompts, cosine similarity, native input limits,
  reranking_hybrid, and all provided candidates are preserved. No sequence
  length, attention, or trust-remote-code override was used.

• Native 1024d is base; supported truncations are 512, 256, 128, 64, and 32.
  All default quantized/rescore conditions remain enabled.

• Model identity, source revision, dtype, dimensions, finite scores, task
  identities, and ranking query coverage were audited in every result file.
  DuckDB independently confirmed all 30 conditions across all 563 tasks.

• No failed or omitted final tasks; batch size remained 4096 throughout.

• The rebuilt viewer rendered Overall, Overall with variants, and MNanoBEIR
  successfully in in-process HTTP checks. Repository validation passed:
  825 tests passed, 4 skipped, with type and lint checks passing.

Checklist

• Result files are committed under hakari-results/hotchpotch__bekko-static-embedding-v1/.
• Result files are compressed .json.xz; no caches, DuckDB files, HTML reports, or local scratch artifacts are included.
• The result JSON records model revision, dataset revision, runtime configuration, and package versions.
• Overall nDCG@10 above was generated from the submitted result files.
• Any non-default prompt, sequence length, attention implementation, candidate ranking, or reranker setting is documented above.