[data-angel]

Repins the atlas-sparkglm lane from RC2 (image 4046c81baa07) to SparkGLM RC4.1.

RC4 (846ee9e1e501, first commit of this PR) is what currently runs on the 01/02 pair; it was deployed with sparkglm-research tools/deploy_rc4.sh. The second commit moves the pin to the RC4.1 hotfix, which is staged on both pair hosts but not yet deployed. It deploys with tools/deploy_rc41.sh.

	RC2	RC4	RC4.1 (this PR)
Image digest	sha256:9a583b07…	sha256:88461891…	sha256:0ddea7cdb372e4407db63dc4f2b5b1a08b4576b8223ded6a4fca3463439ccfb2
Tag / install tree	4046c81baa07	846ee9e1e501	cfb463a952b5 / cfb463a952b517b87baf302b5b3b279a115c5626
Engine (Enntity/atlas)	sparkglm/atlas-20261009-rc2 @ f2b805e7	sparkglm/atlas-20261012-rc4-layered @ 5c523903	sparkglm/atlas-20261012-rc4.1-layered @ 301479a2 (tree ac5a080b)
SparkGLM revision	b6bed259	3e9f91de	35dc1b41 (branch release/rc4, Enntity/sparkglm#46)

RC4.1 fixes long GLM tool calls. The envelope-stuck guard (MAX_TOOL_BODY_TOKENS = 1024) recognized parameter values only by Qwen's <parameter=…> token ids. Every GLM tool call longer than about 1,024 tokens therefore ended with finish_reason "length" and no tool call.

• The fix, ade7b012, tracks value bodies by GLM's <arg_value> / </arg_value> tokens.
• Against RC4's measured tree, RC4.1 adds only that spark-server fix and four CI-only upstream commits. The decode path is unchanged.

RC4 is bit-exact with RC2:

• the logprob hash 246cd1331b41 is unchanged;
• 524/524 probe outputs are byte-identical;
• C4 output equals C1;
• the four-session 204K fill on the deployed recipe scored 12/12.

Single-stream decode is faster: prose 45.6 vs 42.8 tok/s, code 64.9 vs 61.4.

Files: pins.json, the recipe, the backend README, THIRD_PARTY_NOTICES.md, docs/recipes.md, and the recipe test. The changes are the identity substitutions made by sparkglm-research tools/lloom_pin_rc4.py and then tools/lloom_pin_rc41.py.

Checks on RC4.1:

• node test/atlas-sparkglm-recipe.test.mjs passes.
• node backends/atlas-sparkglm/verify-pins.mjs passes.
• node test/smoke.mjs passes.
• The digest resolves anonymously on GHCR.

Before merging:

• Gate first. Deploy RC4.1 and gate it live on the pair.
• Repin after the SparkGLM merge. Once RC4.1: faster prose and code decode, bit-exact with RC2; long GLM tool calls fixed sparkglm#46 merges, replace the SparkGLM revision 35dc1b41c92351d09d47af64e69b44b3393e2334 (and 35dc1b41) with the merge commit in the files above. The install tree, tag and digest do not change. The normal install path pulls by digest and does not need the revision; --source-build does.

🤖 Generated with Claude Code