Reject safetensors vocab shape mismatches

[wang-tong0]

Summary

• add a lightweight safetensors header check after the existing config allowlist validation
• verify embed/lm_head vocab and hidden dimensions match config.json without downloading full shard bodies
• add regression coverage for the UID79-style 151669 vs 151936 vocab mismatch and transient header-read skips

Root cause

UID79 passed the config-based allowlist because config.json advertised a valid Qwen3-32B shape, but SGLang crashed at load time because the safetensors embedding weights had a smaller vocab dimension than config.vocab_size.

Validation

• pytest -q tests/test_model_size_checker.py
• check_model_size on UID79 model now returns pass=False with weight_shape_mismatch:model.embed_tokens.weight.dim0=151669 expected_vocab=151936

Note: the unrelated untracked docs/AFFINE_KEY_FILES_AND_CALL_GRAPH.md file was left out of this PR.