atlas-sparkglm: pin the FP4-Local release, served overlay-free

[data-angel]

Summary

This PR moves LLooM's Atlas SparkGLM lane (linux-nvidia-dgx-spark-2x-glm53-atlas, recipe id unchanged) to the 2026-10-17 FP4-Local release, which is qualified and serving on the production pair:

	Before	After
Enntity/sparkglm	4dee61dc (install tree 2b2970c2f05d)	c28da418 (merge of sparkglm#53, install tree c8ced64316bf)
Engine	Enntity/atlas 289cefbc	sparkglm/atlas-20261017-fp4local @ b3b016ab
Image	ghcr.io/enntity/atlas-sparkglm@sha256:94b837ef…	@sha256:0b91fd790afb2ee3aa5a7f39a0bc5a90db05751436ae7810e00f129c36ebbb7b
Checkpoint	nvidia/GLM-5.3-Flash-NVFP4 @ 423acf37 + per-node overlay	nvidia/GLM-5.3-Flash-FP4-Local @ 76348ddb, served as published

I checked the image against GHCR anonymously: the tag c8ced64316bf resolves to that digest. The config is arm64, with io.enntity.sparkglm.install-tree=c8ced64316bf0b9b2b116d6a0bcdfca52064545b and entrypoint /opt/atlas/serve.py. Release receipts are in Enntity/sparkglm results/2026-10-17-fp4local: prompt-logprob hash 25bc828eea3f, tools 21/21, ToolEvalBench 91.

Overlay-free serving

FP4-Local ships its MTP layer quantized. The release's serve.py loads it directly and refuses an unquantized MTP layer.

• Both members now mount ${modelRoot}/nvidia--GLM-5.3-Flash-FP4-Local read-only as MODEL_PATH.
• Removed: the atlas-overlay mount, setup step convert-atlas-overlay, convert-overlay.sh, and its conversion-headroom.mjs helper.
• No src/ change was needed.

Pin gate (mirrors atlas-sparkqwen)

• New engine pin. install.sh checks that /opt/atlas/source-manifest.json in the image names that commit, in place of the converter checks.
• engine.commit and drafter.revision must be 40-character lowercase git ids. The old gate also accepted uppercase.
• The model must be the pinned FP4-Local revision. The NVFP4 checkpoint needs an overlay this recipe no longer prepares.
• image.entrypoint and image.profilePath must be plain /opt/atlas/... paths, because install.sh interpolates them into the in-image check.
• Unchanged:
  • The digest-only normal path, with no mutable tag pull and no silent fallback.
  • The 64-hex digest check.
  • The image-ID ban.
  • Explicit --source-build.

Docs

Updated the backend README, docs/recipes.md, THIRD_PARTY_NOTICES.md (model license is now the NVIDIA Open Model Agreement; the image still ships the converter, which this recipe never runs), the catalog and index entries, and the CHANGELOG. The "candidate / qualification pending" wording is replaced with the release receipts.

Tests

• node test/atlas-sparkglm-recipe.test.mjs passes. It was updated first and asserts the new pins, the engine-commit contract check, the overlay-free mounts/MODEL_PATH/setup steps, and rejection of the NVFP4 pin, bad engine commits and unsafe in-image paths.
• node test/atlas-sparkqwen-recipe.test.mjs passes.
• npm run interchange:check, npm run package:check and npm run check:shell pass.
• node test/smoke.mjs fails at lloom setup --recipe apple-silicon-qwen36 with "no free backend port in configured range 8234-8234". That failure is identical on a clean origin/main checkout on this machine, and it comes after the recipe-list assertions.

Rollout note

Existing hosts keep a now-unused ${installRoot}/atlas-overlay (about 16 GB) and the old NVFP4 checkpoint. Remove them by hand once the new recipe is applied.

🤖 Generated with Claude Code