fix(cms): stop clamping RegularPath estimates at i32::MAX

[GordonYuanyc]

Problem

CountMin::estimate on the regular path seeds the running minimum with S::Counter::from(i32::MAX):

let mut min = S::Counter::from(i32::MAX);
for r in 0..rows {
    ...
    if v < min { min = v; }
}
min

For counters wider than i32 (i64/i128), any key whose probed cells all exceed 2147483647 never fires the v < min update, so estimate returns the seed unchanged — silently capping every large count at exactly 2147483647.

This bites real workloads: e.g. CMSHeap<Vector2D<i64>, RegularPath> over a SUM(pkt_len) window where true per-key totals run into the tens of billions. Every such key comes back as 2147483647.

Scope

• CMS RegularPath — affected (this fix).
• CMS FastPath — not affected; fast_query_min already seeds from row 0.
• CMSHeap — affected only when built on RegularPath (the struct default Mode); it delegates to cms.estimate.
• CountSketch / CSHeap — not affected; they take a median over f64, no constant seed.

The bug only manifests with counts > i32::MAX and a counter type wider than i32, which is why existing tests (small counts / i32 counters) never caught it.

Fix

Seed the running minimum from row 0's probed cell, then fold in rows 1..rows — mirroring the fast path's fast_query_min. No trait-bound or API changes.

Test

Adds estimate_does_not_clamp_i64_counts_above_i32_max: an i64 sketch loaded with ~35 billion and asserting the estimate is not clamped. Fails on the old seed, passes now. Full countminsketch suite (64 tests) green.

🤖 Generated with Claude Code