Final release: bit-exact KV shard with the disk prefix cache beside it

[data-angel]

Final release: bit-exact KV shard, with the disk prefix cache beside it.

Engine

Enntity/atlas sparkglm/atlas-20261009-final @ b85befc6 = sparkglm/atlas-20261009 (e034500d) plus:

• Disk prefix cache beside the KV shard (feat/kvshard-nvme): each Spark spills and restores only the latents it owns.
• Bit-exact shard (feat/kvshard-exact): few-row MLA attention uses one canonical decomposition whether the shard is on or off, so turning it on changes no output bit.
• Width-invariant rows (fix/canonical-history): one split count for every row count, so a row's bits no longer depend on the DFlash verify width beside it.

Launcher

• SPARKGLM_KV_SHARD=1 no longer refuses the disk prefix cache; .env.example and the tests are updated.
• Image build: compile jobs sized from free memory (ATLAS_BUILD_JOBS overrides). The in-image release checks run in their own Dockerfile stage (ATLAS_IMAGE_CHECKS=0 skips them). A cached rebuild on a Spark takes about 3 minutes.

Checks on the pair (engine b85befc6, dev binary, 4x1m profile)

• Prompt-logprob hash (2,100 rows, two runs per arm): de6a323e9085 with the shard on and off.
• Disk tier beside the shard (pool capped at 7,500 blocks, 8 × 25K conversations, 3 restores each): turns 2–4 are byte-identical to the resident arm, including turn 4's token logprobs. Both ranks show 16 restores, 32,800 spills and no rank disagreement. Restore median 27 µs/block, read 5.2 GB/s.
• Before the width fix, the canonical form's split count followed the verify width (3–11 partitions), and turn 4's logprobs drifted by up to 0.55 after a restore. ATLAS_GLM_KV_CANONICAL=0 isolated it, and a new GPU test (a_row_does_not_depend_on_the_rows_beside_it) fails before the fix and passes after.
• Decode cost (shard on, dev profile), before the width fix: C1 per-step speed within 0.3% of production e034500d; C4 aggregate 68.7 vs 70.1 tok/s, within noise. The fix adds about 10–25 µs per MLA layer at verify widths (microbench), roughly 0.2–0.5 ms per verify step.

🤖 Generated with Claude Code