Add DenseOn and LateOn HAKARI-Bench results

Summary

This PR adds complete standard-scope results for LightOn's paired English
retrieval models:

Both submissions contain all 563 standard evaluation tasks. The Overall values
above use 381 grouped score units derived from the 550 non-overlapping Overall
task results.

NanoBEIR-en validation gate

The official Hugging Face model cards and checkpoint configs were reviewed
before evaluation. Both models were first evaluated on NanoBEIR-en and only
advanced to --all after the scores were consistent with the official 13-task
BEIR profiles.

The official model-card BEIR averages are 56.20 for DenseOn and 57.22 for
LateOn. Nano sampling changes the absolute aggregate, so the per-task
correlations were used as the primary reproduction check.

Runtime and official settings

DenseOn uses the checkpoint's SentenceTransformers configuration: CLS pooling,
768 dimensions, cosine similarity, a 512-token maximum, and the registered
query / document prompts (query: and document: ). The dense submission
contains base, int8, binary, int8_rescore, and binary_rescore for every
task.

LateOn uses PyLate exact MaxSim with 128-dimensional token embeddings, query
length 32, document length 300 (checkpoint model maximum 299), [Q] / [D]
prefixes, and query expansion disabled. It contains one exact-MaxSim base result
per task.

Both checkpoints were pinned by SHA and evaluated in fp32 with explicit SDPA.
No task failed and no batch-size or sequence-length adjustment was made.

Environment: Python 3.12.12, PyTorch 2.9.0, Transformers 5.12.1,
SentenceTransformers 5.4.1, Datasets 4.8.4, CUDA 12.8, and two NVIDIA GeForce
RTX 5090 GPUs (one independent worker per model).

Commands

CUDA_VISIBLE_DEVICES=0 UV_PROJECT_ENVIRONMENT=tmp/.venv_eval_denseon \
uv run hakari-bench evaluate from-model-card \
  --model-card config/model_cards/lightonai__DenseOn.yaml \
  --all --batch-size 64 --device cuda:0

CUDA_VISIBLE_DEVICES=1 UV_PROJECT_ENVIRONMENT=tmp/.venv_eval_lateon \
uv run --group pylate hakari-bench evaluate from-model-card \
  --model-card config/model_cards/lightonai__LateOn.yaml \
  --all --batch-size 32 --device cuda:0

Coverage and plausibility audit

• 1,126/1,126 expected .json.xz files are present and readable.
• Each model has 563 unique task keys, with no missing, extra, or duplicate task.
• All scores are finite and within [0, 1]; all resolved dataset revisions are
  present and internally consistent.
• DenseOn has all five planned variants on every task; LateOn has the planned
  exact-MaxSim base result on every task.
• In the merged leaderboard database, both models have 550/550 Overall tasks.
  Base retrieval means are 0.417095 for DenseOn and 0.446855 for LateOn.
• Relative placement is plausible: DenseOn ranks above BAAI/bge-small-en-v1.5;
  LateOn falls between LightOn's ColBERT-Zero/GTE-ModernColBERT models and the
  smaller AnswerAI ColBERT baseline on Overall.

Checklist

• Results are under hakari-results/lightonai__DenseOn/ and
  hakari-results/lightonai__LateOn/.
• Only compressed result JSON is included; no caches, DuckDB, HTML, or
  scratch artifacts are submitted.
• Result JSON records model/dataset revisions, runtime configuration, and
  package/CUDA versions.
• Official prompts, sequence lengths, attention choice, dtype, variants,
  and late-interaction settings are documented above.