unsloth: drop the pins upstream superseded and repin the conflict-fixed heads - #242

Summary

The last three nightly runs of unsloth-prebuilt.yml failed at "Resolve tag": since upstream merged its own GLM-5-Next and Qwen4Exp MTP, 7 of the 13 pins no longer merge onto a current base tag. This updates pr-set.json and feature-checks.json so the full set merges again on b11368, b11404 and b11417 (the newest tag).

Dropped

What those pins had beyond upstream is not lost:

• Load GLM-5-Next GGUFs converted with the earlier glm5next arch name #239 loads the GLM-5-Next GGUFs already on Hugging Face (glm5next arch, swiglu_limit mmproj) next to the new upstream naming, so nobody has to re-download.
• qwen4exp: run MTP heads that share the target's token_embd and output #240 runs the shared Qwen MTP heads (mtp-*-shared-*.gguf) and the original self-contained heads.
• Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241 carries the CUDA graph cache key, the GLM-5-Next NextN draft graph, the F16 indexer cache under a quantized -ctk, and the macOS half of llama: batched readahead for lazily read gather tables #137's readahead.

On #137 specifically: interleaved cold-cache runs on b11368 (Qwen3.8-Flash-Next UD-IQ1_S, lazy table) show upstream's readahead and #137 at the same throughput and the same major-fault count (95 vs 96, against 10,815 with neither), with byte-identical greedy output.

Repinned to conflict-fixed heads

#70, #95, #149, #152, #157 and #158 are unchanged.

feature-checks.json

• glm5next: owner Load GLM-5-Next GGUFs converted with the earlier glm5next arch name #239, arch probe renamed to upstream's glm5-next.
• qwen4exp-mtp: owner qwen4exp: run MTP heads that share the target's token_embd and output #240.
• llama: batched readahead for lazily read gather tables #137 removed from unchecked; Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241 added there with its reason.

Testing

• Local replay of the workflow's merge loop with additive_merge.py: every pin merges on b11368, b11404 and b11417; Add TML Inkling architecture ggml-org/llama.cpp#25731 and kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes #70 need the additive resolver, as before.
• The b11417 mix builds with CUDA and -DLLAMA_FATAL_WARNINGS=ON. test-llama-archs passes all 374 tests and test-mtmd-impl passes. feature_matrix.py (CPU, as in CI) demonstrates all 7 features.
• Models on the b11417 mix:
  • GLM-5.3-Flash UD-IQ1_S: text and image, with both the original files and the Shard_Rewrite shard 1 + mmproj.
  • Qwen3.8-Flash-Next UD-IQ1_S with the shared and the self-contained MTP heads (89.5% and 90.1% acceptance).
  • Inkling-Small UD-IQ1_S: greedy output byte-identical to the previous nightly; perplexity 41.85, unchanged.
  • DiffusionGemma Q4_K_M: coherent output.
• The pin-set lint steps (schema, every pin checked or knowingly unchecked, script self-tests) pass locally.