ci: extend benchmark gate to G7 — allocation contracts and stdlib-parity ratios

[tr1v3r]

Summary

Extends the benchmark gate from 3 to 7 checks — the library's other performance contracts were unprotected (a Sort/DistinctBy/Take/laziness regression would pass CI silently).

New gates

gate	type	threshold	measured (M3 Pro)	catches
G4 Sort vs direct slices.SortFunc	same-process ratio	≤ 2.5×	1.17–1.22×	adapter re-introduction, O(n²) path
G5 DistinctBy vs Distinct	ratio ×2	≤ 0.6× time, ≤ 5% allocs	0.24–0.27×, 0.3%	fmt.Sprint boxing regression
G6 Take on 100k	deterministic allocs	≤ 50	exactly 6	materialize-all regression (would be ~thousands)
G7 Limit(2).ToSlice() on 100k	deterministic allocs	≤ 200	exactly 12	eager traversal regression

Allocation gates (testing.AllocsPerRun) are hardware-free and zero-flake by construction — the strongest gate kind available.

Documented exclusion

Proposal A6 (multi-section 16+2 speedup) is deliberately NOT a time gate: its speedup is core-count dependent (16 workers need 16 cores to beat 4) and thus cannot be machine-independent. Multi-section correctness stays pinned by TestParallelV2_MidChainParallelReopens.

Verification

• 3 local calibration runs: all gates pass with wide margins (table in file header)
• gofmt/vet/lint 0 issues; full -race suite green; gate runs in ~1.8s