Make installed-wheel timing checks provider-safe

[MateoBodon]

Scope

• replace architecture-specific speedup thresholds in the installed-wheel contract with one bounded 1.50x regression guard
• keep deterministic concurrency and functional wheel checks unchanged
• preserve all pricing code and the separately recorded, artifact-bound v0.4 performance claims

Why

The final wheel matrix passed Linux but observed provider-only timing variance on macOS universal2 x86_64 emulation: a 0.787 ratio narrowly missed a 0.75 benchmark threshold. The preceding Windows run observed the same class of wall-clock variance. A packaging contract should detect catastrophic regressions; it should not serve as the benchmark evaluator.

Verification

• full installed-wheel contract passed locally
• focused concurrency evaluator passed 20/20 repeated local replays
• clean branch is based directly on final main
• remote branch file hash and full tree exactly match the locally verified result