docs: the MLX drafting retention reproduced and bounded; gemma4 resize/#28954 record; parked upstream draft

[glennneuber]

Record of the 2026-09-18 investigation that followed the v0.34.1 deploy — docs only, no code.

• Task doc, "The drafting retention, reproduced and bounded": the deterministic reproducer (an image + a stop-terminated answer + speculative decoding; grammar irrelevant; 14/14 on a size ladder, 0/38 length-terminated), the exclusion list — each item excluded by an instrumented Go-only swap (live-array registry by owning scope, per-bucket byte accounting, creation/close counters by schedule owner), not by reading — with the one n = 1 error corrected in place, and the two-component finding: A, MLX-tracked and owned by nothing on the Go side, the only unbounded growth, removed by no-drafting; B, CUDA-internal and outside MLX's allocator, driven by input-shape variety, present without drafting, bounded (peaks +6.6 GiB early, ~4 GiB at steady state over 60 shapes; the trie fills to its 8 GiB budget as in MLX runner: ~0.147 GiB resident growth per request at fixed context, plateauing at ~28.5 GiB (0.32.15) ollama/ollama#17924).
• Task doc, gemma4 sections: the bilinear→bicubic resize entered with the 0.33.1 fold (llama.cpp #27594) unflagged; the veto on the e2b/e4b attribution closed by a repeat on the deployed 0.34.0 image (identical cell for cell), so every gemma4 move in this fold is #28335's, whose E2B/E4B carve-out is the mechanism; llama.cpp #28954 (the same commit's abort above ~1.2 Mpx) does not reach the fork because the launcher passes -b N -ub N, at the cost of a split non-causal decode at the top rung — with the measurement that decides a fix.
• Retirement register: rows for the launcher's batch vs. the gemma4 image chunk, and for the drafting knob as memory mitigation.
• upstream-mlx-speculation-image-stop-retention.md — PARKED: the ollama issue (reproducer, exclusions, evidence quality) and a note for ml-explore/mlx (the CUDA-internal accounting gap), held per the standing rule until a fix can be proposed.

Instrumented images sync-0.34.1-instr…instr6 and the wt-instr worktree stay on the host for re-runs; nothing of this is deployed.

🤖 Generated with Claude Code