Problem

count_tokens_never_panics called TokenCounter::new() inside the proptest body. Each call to TokenCounter::new() parses the tiktoken cl100k_base BPE vocabulary (~1 MB, ~100k HashMap entries). With 256 proptest cases, this resulted in 256 full BPE HashMap constructions per test run instead of one.

The cl100k_base singleton exists in production code exactly to avoid this, but the test bypassed it.

Fix

Use a LazyLock static so the TokenCounter is initialized once per test binary execution — matching how the production agent uses it (one counter per session).

// before
fn count_tokens_never_panics(s in ".*") {
    let counter = crate::token_counter::TokenCounter::new(); // × 256
    let _ = counter.count_tokens(&s);
}

// after
static COUNTER: LazyLock<TokenCounter> = LazyLock::new(TokenCounter::new); // × 1
fn count_tokens_never_panics(s in ".*") {
    let _ = COUNTER.count_tokens(&s);
}

Result

Test runtime: ~several seconds → 0.204s (measured locally).

Bonus

Add slow-timeout = { period = "30s" } warning override to the ci-partition nextest profile as a safety net — nextest prints a warning if any test exceeds 30s without failing the run, surfacing future regressions early.