Keep managed baseline loops ahead of slower optimizer regions

[chrisbbreuer]

Problem

After warmup crossed the optimizer threshold, the generic arithmetic optimizer superseded a much faster whole-loop baseline artifact and slowed direct samples from about 90 ms to 1.83 s.

Acceptance

• A ready managed baseline loop prevents publication/selection of a generic optimizer artifact for the same chunk in production.
• Optimizer-only effect/property regions and non-loop optimizer entries remain available.
• Focused normal and TSan tests prove the managed baseline remains selected after the optimizer threshold.
• Ten-warm-call direct arithmetic returns to the pre-optimizer scale with stable samples.
• The complete clean-head zig-js/JSC matrix is regenerated and published.

Result

df86d68 retains managed whole-loop baseline artifacts in production; fd0194b scopes that selection policy per interpreter so optimizer-specific unit tests still exercise optimizer publication and execution. Direct arithmetic measures 85.522–85.943 ms across seven focused samples (about 0.19% RSD), and the clean matrix records 85.612 ms. The 64-test optimizer filter and production-policy test pass normally and under TSan; the complete suite passes 1,404/1,404.

[chrisbbreuer]

Completed by df86d68 with the selector correction in fd0194b. Production retains the managed baseline, optimizer tests retain full tier coverage, and all focused/TSan/full gates pass.