bench: 2026-07-06 M1 Ultra full refresh (post VLM-port and fix batches)

[inureyes]

Summary

Full text (169 models) and VLM re-benchmark on mlxcel 0.3.3 (M1 Ultra), covering the five-family VLM port batch (#660-#664), the fix batch (#666-#668), and the qwen3.5-MoE sanitize fix (#671).

Data added

• benchmarks/metal_m1ultra_2026-07-06.csv (text, 169 rows) and benchmarks/metal_m1ultra_vlm_2026-07-06.csv (VLM), plus the two post-fix(qwen3.5-moe): sanitize toward the gate/up/down expert names SparseMoeBlock actually loads #671 single-model re-bench CSVs (qwen3.5/3.6-35b-a3b, back at 82.25 / 78.84 tok/s).
• Per-model mlx-lm / mlx-vlm baselines for 21 newly added models (13 measured, 8 python-side FAIL:warmup), under pylm_m1ultra_*2026-07-06*_single_*.csv.
• Previously uncommitted June baseline CSVs (2026-06-13/14 singles, 2026-06-16 repfresh pair) for a complete audit trail.

Findings

• No decode regressions vs 2026-06-15; the qwen VLM family and small text models run 10-16% faster from the intervening fused-MoE and related work.
• The sweep caught issue fix(qwen3.5-moe): sanitize renames stacked experts to w1/w2/w3 but SparseMoeBlock now loads gate_proj/up_proj/down_proj (load regression since #588) #670 (qwen3.5/3.6-MoE silently failing to load since fix(qwen3-next): make real checkpoints load and run (MoE naming, gated-delta slice, mask, stateful decode) #588's expert-projection rename), fixed the same day in fix(qwen3.5-moe): sanitize toward the gate/up/down expert names SparseMoeBlock actually loads #671.
• New-model baseline highlights: ernie-4.5-vl 160% of mlx-vlm decode, deepseek-vl2 112%, idefics3 119%, qwen3-next-80b 118% of mlx-lm, phi-3-small aq4_64 380%; mistral-small-4-119b loads only under mlxcel. hunyuanocr / qwen3-omni image-mode ratios are marked as first-token-dominated artifacts (10- and 2-token answers); qwen3-omni's text decode (82.81) is at parity with qwen3-vl-30b (81.80).

Docs

model_tests_m1ultra.md: 110 refreshed rows, 14 new VLM rows, 9 new text rows, environment and run-over-run notes for the 2026-07-06 sweep. model_tests.md: index and CSV registry updated. Also updates the ernie4_5_moe_vl / hunyuan_vl parity tests to the lowercase local model-directory names so the suites keep finding their checkpoints.