Feature: hardware-aware SIMD compression execution plans

[awksedgreep]

Parent: #28. Depends on benchmark evidence and the persisted profile contract.

Summary

Map a requested compression profile onto tested effective plans for supported CPU and resource classes, including x86-64 SIMD and ARM64 NEON acceleration.

Scope

• Record baseline x86-64, x86-64 AVX2 (and AVX-512 only if dependencies materially use it), and ARM64 NEON capabilities.
• Define supported memory, optimizer-thread, and CPU-budget classes.
• Verify which codecs and dependencies actually benefit from each instruction set rather than assuming SIMD gains.
• Auto-detect safely, with explicit overrides for containers, throttled environments, and reproducible benchmarks.
• Expose detected capabilities and selected effective plan through public stats.

Guardrails

• Detection selects only a measured plan; it never synthesizes an untested configuration.
• Baseline artifacts remain safe on CPUs without optional instructions.
• Compression output may vary by encoder implementation, but decoded values and compatibility do not.
• Memory and thread limits are first-class; CPU ISA alone does not imply that compact is safe.

Acceptance criteria

• Artifact-platform tests prove safe startup and correct plan selection.
• Published envelopes cover every advertised hardware/resource plan.
• Overrides, fallbacks, and unsupported combinations fail predictably.