The NextN head is published as its own GGUF: nextn_predict_layers and the
trailing blk.<n_layer> block are present, the trunk is not. load_arch_tensors
still created every trunk layer as required, so such a file failed immediately
on `blk.0.hc_attn_norm.weight`. Probe a trunk tensor and relax the trunk to
TENSOR_NOT_REQUIRED when it is absent, and probe the head to relax the MTP
block for the reverse case (nextn metadata kept, head stripped), matching what
deepseek2, qwen35moe and step35 already do. The block index itself was never
the issue: blk.<block_count - nextn_predict_layers> is exactly what upstream
expects.

Also accept the head's output mixer under the model-level hc_head_* names.
It is unindexed in the checkpoint (mtp.hyper_connection_mixer.*), and a
detached head has no trunk mixer to collide with, so published exports write
it there instead of to blk.%d.nextn.hc_head_*.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>