[solderzzc]

What

SL-51, first slice. Wires 6 of the 71 VLM overlay implementations into VLMModelFactory, and adds a processor shim so models without a ported vision or audio tower can load through the VLM factory.

• Shim: SharpAITextOnlyProcessor renders the chat template and tokenizes. It throws mediaNotSupported for images, video or audio instead of dropping them. SharpAITextOnlyProcessorLoadingResolver forces this processor for the text-only model_types whatever processor_class a checkpoint declares, and supplies an empty configuration when preprocessor_config.json is absent. It plugs into the existing VLMProcessorLoadingRegistry; no shared loading code changed.
• Wired keys (SharpAIVLMModelOverlays.vlmWiredKeys): interns1_pro, interns2_mobius, kimi_k3, qwen4_exp, step1 (text-only, use the shim) and intern_s2_preview (subclasses upstream Qwen3.5, keeps its own processor).
• Policy: upstream wins, same as SL-50. The merge is two lines in VLMModelFactory.swift plus one resolver entry in VLMProcessorLoadingRegistry.shared.
• Left out on purpose: the 29 model_types whose creator always throws or returns a model that cannot run. A test pins that none of them is registered.
• Conformance tooling: gen_expectations.py and missing-ops/gen_report.py read the new VLM aggregate; expectations and report regenerated (VLM registry keys 22 → 28).
• No change to .github/workflows or Package.swift.

Tests

• SharpAIVLMOverlaysTests: wired keys resolve in VLMTypeRegistry.shared; aggregate holds exactly them; 29 blocked keys stay unregistered; resolver forces the shim for the 5 text-only types and leaves other models alone; shim tokenizes text with no mask and rejects media.
• SharpAIVLMAirgapLoadTests: for each of the 6 keys, a deterministic non-zero synthetic checkpoint (config.json + safetensors, generated locally, nothing downloaded) loads through the real loader and reproduces the logits via prepare. coversEveryWiredKey fails if a wired key has no case.
• Also run locally: SharpAIModelOverlaysTests, SharpAIAirgapLoadTests, ConformanceRegistryTests, ConformanceMissingOpsTests, ConformanceSwigluGoldenTests, VLMProcessorLoadingRegistryTests, VLMRegistryTests, VLMTextOnlyContinuationTests. Test process EXIT=0.

Not covered

• The air-gap test checks the model creator and loader path. It does not run VLMModelFactory._load end to end (that needs a tokenizer loader), and the synthetic file uses the model's own parameter names, so it does not prove the sanitizers match a real checkpoint.
• None of the 6 has been loaded from a real checkpoint. Real processor_class values for these models are unverified; the resolver overrides them on purpose.

Scouting report and the plan for the remaining PRs: docs/research/sl51-vlm-scout-2026-11-01.md (kept out of the repo).

🤖 Generated with Claude Code