load: one backbone loader, so a vision-language model stops being unr…

[WestWaters]

…eachable

Eleven tools called AutoModelForCausalLM.from_pretrained directly. That auto class REFUSES a vision-language config outright -- "Unrecognized configuration class Qwen2VLConfig for this kind of AutoModel" -- so every one of them died on a VL model, even though nothing they do afterwards cares. Probing, smoothing, abliterating, KL and low-bit all work on the text stack, which a VL model keeps under model.language_model.

It reads as an unsupported model rather than an unasked question, and it put a whole quant ladder in the ditch: pollard-probe could not measure sensitivity on Qwen2-VL, so the build produced one rung and stopped.

pollard_load.load_backbone tries the causal LM, falls back to AutoModelForImageTextToText and AutoModelForVision2Seq, and reports ALL the failures rather than only the last -- a fallback chain that hides the first error is its own debugging problem. text_layers() finds the decoder stack for either family, so tools that count or iterate layers get the same answer on both.

Every tool now uses it: probe, abliterate, gptq, hf_smooth, kl, lowbit, palette, doctor, refcheck. pollard_flybrain delegates to it too, so there is ONE implementation rather than two that drift.

Verified against the model that broke the ladder:

VL model loaded: Qwen2VLForConditionalGeneration
text layers found: 28

A test walks every pollard_*.py and fails if any of them loads a backbone directly again, because a second copy is exactly how this comes back.

56/56 pass.