preflight: measure the mlx-metal-0-34-0 profile on the v0.34.1 fold

[glennneuber]

The v0.34.0 + v0.34.1 folds (#297, #302) moved both payload pins — llama.cpp b10630 → b10864, MLX c793734e → d9add9d1 — so per ADR 0011/0032 Apple Silicon is gated on an unknown (platform, version) until a profile exists. This measures one.

Measured, not derived

On 0.34.0-maxusai-8a7ba949 against :11437 with measure_ladder.py, one freshly restarted server per arch so a log window can only hold its own model.

arch	model	ladder	vs 0.33.2
gemma4	gemma4:31b-mlx-bf16	[1102 × 5] @ prefix 19	identical
gemma4_unified	gemma4:12b-nvfp4	[1102 × 5] @ prefix 19	identical
qwen35moe	qwen3.6:35b-a3b-nvfp4	[68, 146, 578, 2306, 5186] @ prefix 13	identical
qwen35	qwen3.8:27b-nvfp4	[68, 146, 578, 2306, 5186] @ prefix 13	identical

24 of 24 compared values reproduce — ladders, strides, scaling, tolerance, prefixes. Third consecutive payload move to leave token accounting inert here. Generator lines were substituted by script; no measured value was retyped (ADR 0012 rule 8).

Preflight

PASS=19 SKIP=12, no failures, no errors. mlx_payload_pin passes live on a Metal profile for the first time, matching d9add9d against the runner's engine-init line — the check catching binary/payload skew that no version string can express.

The four payload_proof checks skip as designed (budgets_observed = false): the native MLX path emits no load_hparams line.

Invocation note for the next person: preflight on native macOS takes --log-cmd and no --container. Passing --container native makes payload_proof bypass its budgets_observed skip and go_patch_marker shell out to docker. That flag belongs to measure_ladder.py, not here.

Two things recorded honestly rather than copied forward

The stamp is 0.34.0 though the fold is 0.34.1. build-macos.sh derives the base from the newest reachable release tag, still v0.34.0-dynres. So version_pattern admits any 0.34.0-maxusai build; mlx_build identifies the fold and mlx_payload_pin enforces it. The earlier 0.34.0-maxusai-907deffd build (MLX ce916dbb, never deployed) matches the pattern and correctly fails the pin.

The carried budget/pixel fields get a narrower justification than 0.33.2's. That profile carried them on the whole of llm/llama_server.go being byte-identical. Across this fold that is false — the file moved 83/64 and llm/ gained model_split.go. The true claim, and the one in the notes, is that ImageTokensForSize itself is byte-for-byte identical between 2b95b4a5 and 8a7ba949 (32 lines, empty diff) and TestImageTokensForSize passes on this build. budgets_observed = false records that they are carried, not read.

aspect_ladder and the think_format probes are carried from 0.33.2, but not silently: preflight runs both live, and both passed.

Not included

No BINARIES.md row — this build is not deployed. If the :11435 swap goes ahead it needs archiving and a row at that point.

🤖 Generated with Claude Code

[glennneuber]

Follow-up: the build this profile describes has now been deployed and exercised by a full quality campaign.

Deployed to :11435 at 13:17 on 2026-09-18 and verified by three things rather than the version string alone — a live MLX load reporting 0.32.2-61-gd9add9d, and the dlopen path resolving the repo payload, so the Go and MLX halves are the pairing this profile was measured on.

Campaign result (n=1 per cell, the quality setting):

• Think-off, all five nvfp4 models: no quality regression. Every OCR tier, box/label/colour count, invoice field and multi-image verdict identical to 0.33.2-maxusai-2b95b4a5. Answer-token counts identical too. Only scene bbox IoU moved, in the third decimal, and upward on four of five models.
• Think-on: one cell moved — gemma4:31b-nvfp4 drops a 9px fine-text OCR tier. Tracked separately in docs(vision): SOLVED — the 0.34.0 gemma4 vision shift is upstream MLX #3912 #310; it does not bear on this profile's correctness.

The campaign is an independent measurement of the same build and did not disturb any ladder value recorded here.

Also worth noting for reviewers: the native gates that were missing when this was opened have since been run on the same build — x/mlxrunner and all subpackages green, and vision golden parity passing on 31b/26b/12b. See #307 for the golden numbers.