falsification_tests asserts wall-clock speedup/throughput inside a required check — debug build on a contended runner #3359

crates/aprender-serve/tests/falsification_tests.rs asserts SIMD-vs-scalar speedup and a tok/s floor. Both are wall-clock assertions, and they run inside a required check.

Measured

aprender#3114, run 35089783289, job 104773107023, step "Quick tier: lib + integration tests of the selected crates (BSE-17)":

• the job printed quick tier crates: apr-cli aprender-serve;
• that step runs cargo nextest run --profile ci --lib --tests$pkgs — --tests arms every integration target of the selected crates, so this file is dark on main and lights up only when crate selection happens to pick aprender-serve;
• there is no --release, while the file's own header says "Run with: cargo test --test falsification_tests --release -- --nocapture". A scalar-vs-SIMD speedup assertion compiled in debug compares two unoptimised implementations;
• readings inside that single job: size=512 scalar=6.657µs simd=16.776µs speedup=0.40x, and elsewhere size=128 speedup=6.41x. That spread is contention plus debug codegen, not a SIMD regression.

git show origin/main:...falsification_tests.rs | sha256sum equals the same file at #3114's head (da20e1137147deef), so this is a main-level defect. #3114 merely selected the crate that arms it.

Why it matters

This is the fifth instance of "no wall-clock assertion in a required check" in this repo — the ci.yml comment eight lines above that very step already cites the rule.

Fix

Keep the hypotheses and keep printing the numbers; the timing claims must not fail a required check, and must still be enforced where they are meaningful. #[ignore] is banned here for flakes. Numerical/behavioural assertions in the same file stay exactly as strict.

Also note falsify_h5_minimum_throughput (">= 1.0 tok/s on TinyLlama-1.1B Q4_0") is a second throughput assertion in the same file.