[hotchpotch]

Additional public-weight vision decision models need their native rendering, calibration, image ordering, and candidate limits preserved when evaluated through S1MB. This adds native runtime bridges and CLI integration, extends image support in existing bridges, and implements CUDA-only placement with explicit activation memory reserves. Full authored candidates, Noul criteria, and numeric Score levels remain intact; single-image and native candidate-capacity limits remain explicit failures or unsupported scope.

The change preserves tied embedding aliases during meta placement, supports the native Transformers 5.8 kernel registry required by StartLux, adds Solomon int8 handling, and adds the native Aplomb adapter with release calibration, reversed-choice debiasing, FP32 readout and ordered image media. It fixes Winnow projector/source digest accounting and adds a regression test. The training-source review covers 38 measured vision model releases and adds 45 supported benchmark-source disclosures, preserving unknowns and distinguishing source-family exposure from evaluation-example use. No datasets, weights or measurements are included in this source PR.

Validation at a813909: uv run --no-sync tox in evaluator/ passed with 675 tests passed and 18 skipped; Ruff and type checks passed. GitHub evaluator and viewer checks passed. StartLux, Rune and Xor-26B recovery runs completed English, Japanese and Vision with zero failed decisions. Complete measurements for 14 models, including the corresponding training disclosures, are submitted separately in https://huggingface.co/datasets/hotchpotch/s1mb-result/discussions/108; all three mandatory completion gates passed.

Remaining limits: Clef-Omni exceeds the available reserved CUDA memory and has no completed scope from these attempts. Solomon remains incomplete. These paths are not claimed to have full GPU qualification or leaderboard eligibility. CPU/meta and synthetic checks do not establish GPU support for every model or capacity. No incomplete measurements are part of the results submission. Unsupported Box is optional and is not fabricated. No hosted inference APIs were used.

[hotchpotch]

@hotchpotch Thank you for the native vision adapters and the detailed qualification notes. Review of head 7e6133cc73d2bbcf5c894e39b04356d64522e18f found one provenance defect to fix before merging:

P2 — Preserve the Python source fingerprint in the Winnow vision path. In evaluator/src/s1mb/adapters/winnow.py:131, the projector hash overwrites digest, which originally contains the Python source hash from source_path(source) at line 97. Line 158 then records the projector hash as source_python_sha256. Use a separate projector-digest variable and add a focused check that the source and projector fingerprints retain their distinct meanings.

Latest-head evaluator/viewer CI is green, and 80 focused synthetic tests passed locally in an isolated, credential-free environment. Those checks do not qualify the pending StartLux/Clef-Omni GPU paths or Rune/Xor long-input recovery described in the PR. Please retain draft status until those declared readiness items and the fingerprint fix are complete, then mark the PR ready for review. No GPU inference or model-repository program execution was performed for this review. Thank you!