Flaky timer_precision on loaded CI hosts (Windows 756µs; WASI 0.3 2.31ms) #30

timer_precision (crates/turnloop-contract/src/native_surface.rs) failed on windows-2025 default mode in main CI run 34947726746 (commit 7e367cd):

timer precision exceeded platform bound: median lateness 756.2µs

It passed on the same code in the PR run minutes earlier, and on every other recent run. A failing main CI also skips the release workflow, which is how this blocked publishing 0.1.0-alpha.3 until the job was re-run.

The same test also failed once on WASI 0.3 during #22 (median 2.31 ms against a 2 ms bound, host load average 150–180), and passed on rerun.

Options to consider, without weakening the bound for a quiet host:

• measure the platform's achievable precision once at test start (a short calibration) and compare against that, failing only on a real regression;
• report the distribution (p50/p90/max) and require the median only, with a documented allowance when the host is demonstrably loaded;
• retry a bounded number of times and fail only if the median stays out of bound.

Whatever is chosen must still fail when the backend genuinely loses timer precision (prove it by mutation).