Feat/qwen3.6

[yogthos]

Registry: new xlarge-tier Qwen3.6-27B-MTP-UD-Q4_K_XL entry with no-artifacts lens status and download_url=None (manual download).

Inference: bump llama.cpp to 25558268 (MTP merge, PR #22673). Regenerate PC-202 expose-hidden-states.patch for the new upstream where server-context.cpp gained cstddef/cinttypes includes and renamed ctx→ctx_tgt / model→model_tgt.

macOS launcher: gate --spec-type draft-mtp --spec-draft-n-max 2 behind ATLAS_ENABLE_MTP=1 in atlas-llama-macos.sh. MTP is default off — net throughput loss on Apple Metal (llama.cpp #23752).

Lens: train cost_field.pt for 5120-dim Qwen3.6-27B residuals (200 epochs, contrastive ranking loss). Metric tensor is a placeholder — G(x) scoring uses XGBoost, not the legacy tensor. Doctor lens_weights check still requires metric_tensor.pt on disk.

[itigges22]

Thanks for the Qwen3.6-27B work! Before we can land the model in the registry so others can use it, we need the Hugging Face links for the artifacts — the entry is currently incomplete:

1. Model GGUF download_url — it's None (manual download) right now, so the catalog can't fetch it and atlas model install won't work for anyone else.
2. Lens + ASA artifact links — the entry shows no-artifacts lens status. We need the published Lens (cost_field / gx) and the ASA control vector on HF so the registry can point at them and the model works out of the box.

Good news: the lens/ASA publish pipeline is working now — the gemma-4-12b registry PR (#123) went through cleanly, so there were fixes that should make this smooth for you. The full end-to-end flow is documented here: https://github.com/itigges22/ATLAS/blob/main/docs/PUBLISHING.md — atlas publish does Lens + ASA + the registry PR in one shot (see "Publishing everything at once").

Once the artifacts are uploaded and the registry entry has the real URLs, I can add it to the registry / main dataset and validate it end-to-end. Appreciate it!

[yogthos]

ASA control vector for Qwen3.6-27B-MTP-UD-Q4_K_XL

Uploaded to https://huggingface.co/yogthos/atlas-asa-qwen36-27b-mtp-ud-q4_k_xl

• File: ast_edit_steering.gguf
• SHA256: 3e6221f01e2fdf7a6d46e73d7103729933cba41b822471f6e1a5fef7c4c3fc77
• License: apache-2.0
• Layer: 27

[yogthos]

Lens artifacts for Qwen3.6-27B-MTP-UD-Q4_K_XL

Uploaded to https://huggingface.co/yogthos/atlas-lens-qwen36-27b-mtp-ud-q4_k_xl

• Files: cost_field.pt, metric_tensor.pt
• cost_field.pt SHA256: 5f9b707ee0c75278e54c0fc415175dbd1477919e64744611a685ebbb353881dc
• License: apache-2.0

[yogthos]

These ASA and lens should work.

[yogthos]

Rebased on main (the gemma-4-12b registry entry from #123 conflicted with the Qwen3.6 entry — both kept, registry is 8 entries now) and pushed an update. Status against the asks:

• GGUF `download_url`: already populated — the entry points at `unsloth/Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-UD-Q4_K_XL.gguf` with a pinned SHA256, no HF token needed. atlas model install works.
• ASA control vector: published — https://huggingface.co/yogthos/atlas-asa-qwen3.6-27b-mtp-ud-q4_k_xl (`ast_edit_steering.gguf`, layer 27, 5120-dim). Entry now carries `asa_status=supported` + `asa_hf_repo` + `asa_artifact_files`.
• Lens: still `no-artifacts`. The C(x) cost_field is trained but the G(x) classifier half (`gx_*.json`) was never trained, so atlas lens publish would only ship half a lens and G(x) would no-op. Ill train G(x) and publish the lens in a follow-up rather than land a partial one here.

Also dropped an empty 0-byte `metric_tensor.pt` that had been committed by accident. So this PR lands the model + ASA; the lens follows once G(x) is trained.

[itigges22]

Marking blocked pending rework — dev's artifact contract moved under this PR (it's also in conflict). What a mergeable version needs:

1. Rebase on dev (flow is dev → staging → main). The llama.cpp pin bump + regenerated patch should be its own PR — it changes the default runtime image for every model, and dev is now pinned to a different rev anyway.
2. No binaries in git: cost_field.pt goes to an HF repo via atlas lens publish (which also writes the required model_identity.json), with lens_hf_repo + per-file lens_artifact_sha256 in the registry entry — see docs/PUBLISHING.md.
3. Complete the bundle: current requirements are model_identity.json, cx_normalization.json, gx_xgboost.json, gx_weights.json, gx_thresholds.json (the doctor/lens-check contract). The onboarding loop that produces them: atlas bench → atlas lens build → atlas asa build → atlas publish. Note the metric-tensor path was removed on dev, so the placeholder-tensor question is moot.
4. ASA hash pin: asa_artifact_sha256 for the published vector.
5. Hardware evidence beyond macOS/Metal for an xlarge CUDA-tier entry (registry contract tests now gate entries; SUPPORT_MATRIX.md levels apply).

The registry entry itself (pinned URL + sha256 + honest no-artifacts lens status) is in good shape — most of the work is publishing the bundle through the supported flow.