Engine sparkglm/atlas-20261009: copy drafts on, streamed GLM text keeps tool/user/assistant, opt-in KV shard

[data-angel]

Pins Enntity/atlas sparkglm/atlas-20261009 @ e034500d (the 10-07 engine plus 9 cherry-picks; the generic ones are also on upstream/glm53-flash-20261009).

Fixes

• Streamed GLM replies lost the words "tool", "user" and "assistant." A Qwen-only role-leak filter ran for every model and dropped any streamed delta that was just one of those words; GLM streams " tool" as one token. In the MLPerf BFCL outputs, SparkGLM's streamed text had "tool" 15 times against 156 for another GLM-5.3-Flash server ("I don't have a that can ..."). Non-streaming was fine. The filter now runs only for ChatML tokenizers.
• Poolside tool-call headers are held until the call validates (sparkqwen/rc15-fixes 25d14fcb): streaming no longer announces calls the blocking path would drop (unregistered names, truncated calls); clears the two glm_tool_boundary test failures.

Speed

• Copy (prompt-lookup) drafts on by default, COPY_REPLY_MATCH=16, COPY_MISS_MAX=3. On the pair, same binary on/off: full-file edits 63.0 / 61.9 -> 79.5 / 77.5 tok/s (+26%); RigMark code / prose / structured 66.5 / 36.0 / 92.6 -> 66.3 / 36.6 / 91.7; greedy text byte-identical with the verify width pinned; quality probe 40/40 arithmetic + 12/12 needles on and off.
• KV shard, opt-in (KV_SHARD=1, start-node.sh --kv-shard): each Spark stores half the latents. At 0.89 with the carveout: 120,186 blocks (~1.92M tokens) against 67,608, idle free memory 8.4 GB on both arms, C1 decode 42.9 vs 42.0 tok/s, C2 54.5 vs 54.1, C4 69.8 vs 71.8. Cannot be combined with PREFIX_CACHE_DIR (refused at launch).

Release checks and the image tag follow in this PR.

🤖 Generated with Claude Code

[data-angel]

Release checks, image bcde3c97e22c (engine e034500d; launcher defaults: KV shard on, 4x1m, 0.91, carveout, no disk tier):

• logprob hash a26c76e6ac39 (reproducible). It differs from 8c75d2886794 because the shard merges split attention in FP32: over 48,875 prompt tokens the sum moves -38741.28 -> -38741.50. With the shard off (pin9 check, same engine) the hash is 8c75d2886794.
• pool with the shard at 0.91 + carveout (clean boot): 163,680 blocks, ~2.62M tokens (1.36M unsplit); four-session 204K fill 12/12 + 8/8, rank 0 at 4.53 GB free at worst, rank 1 5.50 GB.
• 4x1m: a 925,096-token needle prompt answered correctly (prefill 486 s), 5.05 GB free at worst.
• copy drafts: edits +26%, RigMark unchanged, byte-identical at a pinned width, quality probe 40/40 + 12/12 on and off.
• pin9 (shard off): vision 6/6, decode C1 ~43 tok/s, prefill 32K 12.36 s / 64K 23.65 s, 0 errors.
• tool arguments keep string types; streamed text keeps "tool" / "user".
• image build + start at 10 jobs: 180 s.
  Image pushed to GHCR.