Report Apple Metal compatibility

[SamJSui]

What changed

• publish the first Apple M3 Pro / Metal correctness and performance matrix
• record Massively 0.96's Metal pipeline-layout failures without inventing timings
• make the POSIX comparison runner continue after per-case failures, matching PowerShell
• condense the README from 421 to 217 lines and lead with current Massively, Apple Metal, and wgpu_sort benchmarks

Why

wgpu-primitives passes all 64 release tests and every benchmark validator through 100M items on Apple Metal. Pinned Massively 0.96 cannot create its compared pipelines: generated layouts request 42 or 47 compute-stage storage buffers while the adapter limit is 29. The old POSIX harness failed fast, which discarded valid wgpu-primitives measurements and hid the compatibility boundary.

The README now presents the current competitive evidence first and links to detailed reports instead of duplicating their history.

User impact

The comparison artifact now distinguishes successful timings from unsupported implementations. Apple users get a reproducible absolute Metal baseline; failed competitor cells remain explicit and are excluded from speedup calculations. New users can see the cross-library results, install the crate, and reach an end-to-end example without navigating a 421-line benchmark narrative.

Upstream tracking

• Massively defect: WGPU primitives fail pipeline creation on Apple Metal due to storage-buffer binding limits massively-labs/massively#62

Validation

• cargo fmt --all --check
• cargo clippy --all-targets --all-features -- -D warnings
• cargo test --release --all-targets - 64 tests passed on RTX
• cargo package --allow-dirty
• all relative README links resolve locally
• python3 benchmarks/massively-comparison/test_aggregate.py - 3 tests passed on M3 Pro
• sh -n benchmarks/massively-comparison/run.sh
• clean canonical M3 Pro matrix - 36 successful wgpu-primitives runs, 36 recorded Massively failures, 12 aggregates, 0 fabricated comparisons